WorksComputer Science
ShadowTransfer: A Geographic Transfer Benchmark for Overhead Shadow Detection
Abstract
Overhead aerial imagery feeds dense-prediction pipelines that are routinely trained in one geographic region and deployed in another, yet no benchmark measures how shadow detection models transfer across regions. Existing overhead shadow datasets either lack a geographic transfer protocol or rely on automatically generated labels. We introduce ShadowTransfer, a benchmark of 4,500 human-verified shadow masks across three U.S. cities at two native resolutions, paired with a leave-one-city-out evaluation protocol. Under vanilla training, holding out the test city costs 3.5 to 7.8 mIoU across four architectures, the gap remains significant when the training set grows from 225 to 900 images, and six of eight adaptation methods degrade more than vanilla training on average. We propose a three-probe framework that diagnoses where and how the transfer breaks: the failure operates in feature space, beyond the reach of pixel-histogram matching, shows up as a confidence collapse on shadow pixels while background confidence stays intact, and is recoverable by retraining different architectural components depending on the model: encoder for some, decoder for others. These findings replicate on INRIA building footprints, suggesting they describe geographic transfer beyond shadows. Guided by the diagnosis, we build Spectral Information Bottleneck (SIB), a plug-in module with one component per finding. Across the 9 (architecture, held-out city) cells, SIB has the smallest mean gap to the within-city upper bound of the ten methods we evaluate (6.42 mIoU), level with FDA. The benchmark, LOCO protocol, code, and trained models are released to support reproducible evaluation of geographic transfer in overhead dense prediction.
Introduction
Overhead aerial imagery feeds segmentation pipelines that are deployed across geographic regions. Shadow masks are a core input to several of these pipelines: solar irradiance estimation uses them to separate shaded from sunlit surfaces [1], urban heat mapping uses them to correct surface temperature readings [2, 3], and building footprint extraction uses them to resolve occlusion boundaries [4]. These pipelines are routinely trained on data from one region and deployed in another [5, 6].
Geographic shift is a known failure mode in segmentation models across domains. Medical imaging models underperform across hospitals due to scanner and protocol differences [7], and aerial land-cover models show similar drops on unseen regions [8, 9]. For overhead shadow detection, no comparable evaluation infrastructure exists. AISD, the largest human-verified overhead shadow dataset, contains 514 images and defines no train-test split across cities [10]. S-EO scales further but relies on automatically generated masks with documented alignment artifacts, and again reports no geographic holdout [4]. Other widely used shadow benchmarks (SBU, ISTD) target ground-level imagery and do not match overhead geometry [11, 12]. Table 1 summarizes the landscape. Without a benchmark that pairs human-verified labels with a defined geographic protocol, transfer failure in overhead shadow detection cannot be measured, diagnosed, or addressed.
Table 1. Comparison of shadow detection datasets. No existing overhead dataset combines human-verified labels, multiple cities, multiple native resolutions, and a defined geographic transfer protocol. ShadowTransfer fills this gap. ISTD is included for context only; ground-level geometry is not a substitute for overhead evaluation. “Transfer protocol” = a defined geographic train-test split with formal evaluation guidelines. ✓ = present; × = absent; n/r = not reported.
To bridge this gap, we introduce ShadowTransfer, a benchmark for measuring geographic transfer in overhead shadow detection. It contains 4,500 human-verified shadow masks across three U.S. cities (Chicago, Miami, Phoenix) at two native sensor resolutions (0.3 m and 0.6 m), with a leave-one-city-out (LOCO) protocol that trains on two cities and tests on the third.
Figure 1 shows the size of the gap this protocol exposes: under vanilla training on the high-resolution cells, holding out the test city costs between 3.5 mIoU (CrossEarth) and 7.8 mIoU (OGLANet), and existing domain-adaptation methods help on some architecture-city pairs and fail catastrophically on others. The major contributions of this work are threefold:

Figure 1. SHADOWTRANSFER and the geographic transfer gap. Left: paired test patches from three U.S. cities at two native NAIP resolutions (0.3 m/px, 0.6 m/px); each column shares ground coordinates between rows, and the white box on each 0.6 m/px tile marks the paired 0.3 m/px footprint. Shadow masks are human-verified across six categories (bottom legend). Right: two-class boundary-tolerant mIoU under vanilla training on the high-resolution cells, with each city held out in turn and results averaged over the three rotations (inset). Open marker = within-city upper bound; filled marker = LOCO held-out city; = mean gap; error bars are 95% CIs. Methods: Table 2.
We release ShadowTransfer2 with its LOCO benchmarking and statistical reporting.
We provide a three-probe evaluation framework that characterizes how models fail under geographic transfer: where in the pipeline the shift acts, how the prediction distribution distorts, and which architectural components carry the failure
We present SIB, a case-study module designed from the framework’s findings, with the smallest mean gap to the within-city upper bound of the ten methods we evaluate, level with FDA. Together, the benchmark and framework give practitioners a way to measure geographic transfer risk before deployment and to diagnose where in a model that risk concentrates.
Related Work
Shadow detection in overhead imagery. Overhead shadow detection assigns a binary shadow label to every pixel of an aerial or satellite image. The field was established by AISD [10], a 514-image manually annotated benchmark spanning five cities at 0.3 m resolution. S-EO [4] scales to roughly 19,000 images by deriving masks from digital surface models and sun geometry; the authors document temporal-misalignment artifacts and recommend the dataset for training rather than evaluation. Airborne-Shadow [13] adds a manually annotated overhead set with a heavy/light shadow distinction, but is restricted to German imagery and does not report a geographic split. Recent methods (MAMNet, MBCSANet, CSDFormer, OGLANet) report F1 scores within roughly one point of each other on AISD [14, 15, 16, 17], indicating that the in-domain benchmark has plateaued. Additionally, ground-level shadow datasets [12, 11, 18] are not substitutes for overhead evaluation: viewing angle, and shadow geometry differ. None of these overhead datasets defines a geographic train-test split, and the most recent overhead shadow survey identifies cross-dataset generalization as an open problem [19].
Domain adaptation for dense prediction. A range of methods make dense prediction robust to source-target appearance differences by intervening at one stage of the pipeline. At the input, FDA swaps low-frequency Fourier amplitude between images [20], IIM derives illumination-invariant inputs [21], and BSM mixes styles within a training batch through a frozen AdaIN network [22]. In feature space, ISW whitens photometrically sensitive covariance dimensions [23], SegDeSiC aligns encoder features through a geo-coordinate regression loss [24], MRFP+ randomizes high- and low-frequency feature components [25], and FADA processes the Haar-wavelet bands of foundation-model features separately [26]. At test time, TENT minimises prediction entropy over normalisation-layer parameters [27], and SAR stabilises this update with sharpness-aware minimisation and by excluding high-entropy samples [28]. Each method applies its correction at the same locus for every architecture, and no prior work tests whether that locus is consistent across architectures for overhead dense prediction. Remote-sensing foundation models address the shift through the representation instead: CrossEarth trains adapters and a Mask2Former decoder on a frozen DINOv2 backbone, with style augmentation from an external image pool and masked image modelling [9].
The ShadowTransfer Benchmark
Dataset Design and City Selection
ShadowTransfer contains 4,500 RGB patches of size pixels, sampled from National Agriculture Imagery Program (NAIP) orthoimagery across three U.S. cities: Chicago, Miami, and Phoenix. Each city contributes 1,500 patches, split evenly across two native NAIP resolutions, 0.3 m and 0.6 m ground sampling distance (750 patches per cityresolution pair). The three cities span distinct urban morphologies and climatic regimes, so the leave-one-city-out protocol in § 3.3 measures transfer across genuine appearance shifts rather than nearby variants of the same scene type. Other potential confounders (sensor type, image format, file conventions) are held fixed by the shared NAIP source. We retain the two native resolutions rather than synthetically downsampling 0.3 m imagery, since NAIP releases them as separate native acquisitions with different sensor characteristics; collapsing these differences would understate the gap deployed systems face. Preserving both resolutions also supports future study of resolution shift under the same protocol. Within each city, patches are sampled uniformly inside the administrative boundary, with at most one patch per source NAIP tile to limit spatial autocorrelation. For each city-resolution pair, the 750 patches are split into 450 training, 150 validation, and 150 test images, fixed at the benchmark level and released with the dataset. To support resolution-paired analyses, we additionally co-locate 150 coordinates per city imaged at both resolutions (300 paired patches per city). Sampling code, random seeds, and per-patch metadata (city, resolution, longitude/latitude, source tile ID) are released with the benchmark; within-city split distributional diagnostics are in Appendix § 11.
Annotation Pipeline
Pixel-level shadow masks are produced by 7 trained annotators in a three-phase pipeline that jointly solves throughput and label quality. Phase 1 establishes the label distribution from scratch through human annotation. Phase 2 trains an interim shadow detector on Phase 1 labels and uses its predictions to seed subsequent batches, which annotators edit rather than redraw. Phase 3 is a retroactive quality-control pass that intersects each prior segment with the final-model prediction and returns disagreements for adjudication. Across all phases, every 100-image batch includes 5 gold images with frozen reference masks; Cohen’s on these images monitors annotator drift across batches. Each shadow segment carries one of six type labels (taxonomy in Appendix § 10); the six-way typing is released for downstream analyses, but the LOCO benchmark in § 3.3 evaluates binary masks (shadow vs. background). Tooling, per-phase throughput, inter-annotator agreement statistics, and Phase 3 adjudication rules are in Appendix § 9.
The LOCO Protocol
We define the leave-one-city-out (LOCO) protocol over the city set . For each holdout city , the training set is sampled from the two non-holdout cities at a fixed 50/50 ratio, yielding images (225 per training city). The validation set is images drawn from the same two training cities (150 per city), and the test set is the held-out city’s 150-image test pool. The within-city upper bound for each (city, resolution) is a model trained on that pair’s 450 training images and evaluated on its 150 test images, so the LOCO-vs-upper-bound gap isolates geographic shift from training-set scale. All three holdout rotations are evaluated; results aggregate over rotations and the two native resolutions, giving 6 test cells per (architecture, method) pair. The primary metric is boundary-tolerant mIoU, the mean of shadow and background IoU with a 2-pixel don’t-care band, since shadow boundaries are inherently imprecise to annotate at sub-pixel resolution; full metric definition and excluded-mask handling are in Appendix § 12. Per-cell significance uses paired bootstrap on per-image values ( resamples, two-tailed -value against the within-city upper bound).
Baselines
We evaluate four architectures spanning representative design families: MAMNet [14] (CNN with criss-cross attention), OGLANet [17] (CNN with multi-scale prediction-stage output aggregation), DINOv3 [29] (ViT with self-supervised pretraining, fine-tuned end-to-end), and CrossEarth [9] (remote-sensing foundation model with a frozen DINOv2 backbone and a Mask2Former decoder).
MAMNet, OGLANet and DINOv3 are each paired with nine configurations: Vanilla (source-only, no domain adaptation), FDA [20], SegDeSiC [24], IIM [21], ISW [23], MRFP+ [25], FADA [26], SAR [28], and BSM [22]. Adaptation methods leave the architecture unchanged. The first seven are tested on all (method, architecture, holdout city, resolution) configurations, and SAR, BSM and CrossEarth on the high-resolution cells, for 158 LOCO test cells of 150 images each. Architecture details, pretrained weight sources, and adaptation-method hyperparameters are in Appendix F. Table 2 summarizes results pooled across the three holdouts and two resolutions; per-cell numbers and confidence intervals are in Appendix G.
Table 2. LOCO baseline performance across architectures and adaptation methods. Two-class boundary-tolerant mIoU ( px). For each architecture-method pair we report mIoU pooled across the 6 test cells per pair (3 held-out cities 2 native resolutions, 150 paired test images per cell). Upper is the within-city upper bound (per-architecture, pooled across the same cells). is the mean gap to the upper bound (mIoU points; negative is degradation). Sig. is the count of cells, out of 6, with significant degradation under paired bootstrap (, ). †High-resolution cells only (3 held-out cities); Sig. is given out of the evaluated cells. The Pooled block pools MAMNet, OGLANet and DINOv3.
| Method | LOCO mIoU D | M | O | C† | (mIoU) D | M | O | C† | Sig. D | M | O | C† | Pooled | Sig./18 |
| Upper | 77.6 | 81.7 | 74.9 | 78.8 | — | — | — | — | — | — | — | — | — | — |
| Vanilla | 70.0 | 77.1 | 67.0 | 75.3 | −7.6 | −4.6 | −7.9 | −3.5 | 6 | 6 | 6 | 3/3 | −6.68 | 18 |
| FDA | 69.4 | 77.0 | 67.0 | 75.1 | −8.1 | −4.7 | −8.0 | −3.7 | 6 | 6 | 6 | 3/3 | −6.93 | 18 |
| SegDeSiC | 73.0 | 76.1 | 65.0 | – | −4.6 | −5.6 | −9.9 | – | 5 | 6 | 6 | – | −6.68 | 17 |
| IIM | 71.4 | 71.9 | 65.5 | – | −6.1 | −9.7 | −9.4 | – | 6 | 6 | 6 | – | −8.41 | 18 |
| MRFP+ | 72.7 | 71.4 | 60.9 | – | −4.9 | −10.3 | −14.0 | – | 6 | 6 | 6 | – | −9.74 | 18 |
| FADA | 67.0 | 72.3 | 63.8 | – | −10.6 | −9.4 | −11.1 | – | 6 | 6 | 6 | – | −10.36 | 18 |
| ISW | 59.0 | 68.3 | 63.9 | – | −18.5 | −13.4 | −11.1 | – | 6 | 6 | 6 | – | −14.32 | 18 |
| SAR† | 68.4 | 73.1 | 65.4 | 75.3 | −6.3 | −6.6 | −7.6 | −3.5 | 3/3 | 3/3 | 3/3 | 3/3 | −6.86 | 9/9 |
| BSM† | 66.2 | 72.8 | 62.3 | 75.7 | −5.7 | −6.9 | −10.7 | −3.0 | 2/2 | 3/3 | 3/3 | 3/3 | −8.05 | 8/8 |
| Avg. | — | — | — | — | −8.6 | −8.2 | −10.2 | — | — | — | — | — | — | — |
| # Sig./42 | — | — | — | — | — | — | — | — | 41 | 42 | 42 | — | — | — |
caption: D: DINOv3 M: MAMNet O: OGLANet C: CrossEarth. Avg. and # Sig. cover the first seven methods. –: not run; FADA, IIM, MRFP+, SegDeSiC and ISW are defined on a per-pixel head, which CrossEarth’s query-based Mask2Former decoder lacks. CrossEarth SIB (−4.4, 3/3) is in Table 36.
157 of the 158 test cells degrade significantly under geographic transfer (, paired bootstrap; 151 after Bonferroni correction over the 158-test family), with mean LOCO degradation of mIoU and a worst single cell of mIoU. A data-scaling test indicates that the gap is not a small-data effect: quadrupling the LOCO training set from 225 to 900 images narrows it by 0.80 mIoU (95% CI [0.46, 1.14]), from 6.48 to 5.50, with the upper bound significantly ahead in every cell but one (high resolution, three seeds; Appendix H.1). Two findings in Table 2 shape the rest of the paper. First, the cost of an adaptation method depends on the architecture: ISW costs DINOv3 mIoU against on OGLANet, while MRFP+ costs OGLANet against on DINOv3. Second, no adaptation method is uniformly safe: six of eight adaptation methods degrade more than Vanilla on average (five significantly, Holm ), and every method except SegDeSiC degrades every cell it is evaluated on, on all four architectures. Without access to the target city, Table 2 gives a practitioner no basis for choosing an adaptation method. § 4 asks where in the architecture the failure lives and why uniform corrections fail non-uniformly across them.
Evaluating Geographic Transfer Failure
Framework
Table 2 showed that no adaptation method is uniformly safe across architectures, suggesting the failure is not a single mechanism a single correction can address. Yet the gap is consistent enough within each architecture that we can ask where in the model it lives. We answer this with three controlled interventions on the LOCO models from Table 2, ordered by what a practitioner would try first. § 4.2 tests the cheapest possible fix: aligning pixel statistics across cities, ruling in or out a whole class of preprocessing-based recoveries. § 4.3 then asks whether the model’s outputs are still usable under a confidence threshold, since deployed systems consume thresholded predictions rather than raw probabilities. § 4.4 localizes which component of each architecture carries the failure. Figure 2 grounds what we are explaining: under LOCO, models keep spatial localization but lose foreground-class confidence: sharp upper-bound predictions become washed-out heatmaps over true shadow regions, not hallucinated shadow elsewhere.

Figure 2. Failure mode and recovery under geographic transfer. Rows: three architectures on tiles from their leave-one-city-out (LOCO) rotations, together spanning all three held-out cities. Columns: input, ground-truth mask (white = shadow), and predicted shadow probability under Upper bound (within-city training), LOCO, and SIB (§ 5); shared colormap [0, 1]. : mean predicted shadow probability inside ground-truth shadow. What to look for: LOCO collapses while background stays dark — failure is washed-out predictions over true shadow regions, not hallucinated shadow elsewhere. SIB recovers foreground confidence on the same regions across all three architectures with varying magnitude.
Failure operates in feature space, not pixel space
Test design. If the LOCO failure were a pixel-statistics shortcut, aligning the test image’s pixel histogram to the training-city distribution should at least partially recover the gap. We test this directly. For each (architecture, holdout city) pair under vanilla LOCO, we compute the aggregate intensity histogram (the mean of the three channels) of the source-city training images. We map each test image’s intensity CDF onto this histogram, restore each pixel’s ratio between its channels and its intensity, and run inference with the trained model unchanged.
Result. Histogram matching makes things worse, uniformly. Across the 9 vanilla LOCO cells, mean mIoU drops by 0.45 relative to LOCO itself (paired bootstrap 95% CI [, ], ); not a single cell improves. To put the magnitude on a scale, we use the per-image recovery ratio where means full recovery; means no change from LOCO; means the intervention is worse than doing nothing. Per-architecture medians fall in [, ], with all 95% CIs excluding zero: at the median, matched images lie 2.5 to 3.7 times as far from the upper bound as LOCO itself. The same rule improves none of 27 vanilla LOCO models trained on 225, 450 or 900 images (three architectures, high resolution; Appendix § 15.§ 1).
Training-time appearance augmentation. We also retrain both the LOCO model and its upper bound with appearance-only augmentation (colour, illumination, contrast, blur, noise; geometry preserved) on the 9 high-resolution (architecture, held-out city) cells. Retraining with a different random seed alone moves a cell’s gap: across three vanilla seeds, the standard deviation of the gap ranges from 0.24 to 4.26 mIoU between cells. Before training the augmented models, we therefore set each cell’s threshold at twice this standard deviation and count a gap as changed only if augmentation moves it by more. Six of nine gaps do not change. The mean gap moves from to mIoU, with OGLANet/Phoenix and DINOv3/Phoenix narrowing and DINOv3/Miami widening. Phoenix also has the largest seed variance of the three cities, and three cities cannot separate a Phoenix-specific appearance effect from that variance (Appendix § 15.§ 2).
Implication. Geographic transfer failure is not a pixel-level shortcut. The features the LOCO model relies on encode the joint distribution of intensity, texture, and spatial pattern in source-city imagery. Aligning the marginal pixel histogram alone breaks this coupling without replacing it with anything useful. This rules out a whole class of cheap fixes a practitioner might try first (like gamma correction, contrast normalization, single-channel histogram standardization) and tells us any successful intervention has to operate at the feature level.
Class-asymmetric selective-prediction failure
Test design. A practitioner deploying segmentation under domain shift consumes confidence-thresholded predictions, not raw outputs. The relevant question is therefore not the average error but how the threshold-conditioned error rate degrades under shift. The selective-prediction (SP) framework [30, 31, 32] measures threshold-conditioned error: per image, sort pixels by confidence and compute the cumulative error rate at each coverage level . The Area Under the Risk-Coverage curve (AURC) integrates this curve. The SP gap = AURC − AURC measures how much threshold-conditioned error rises under transfer; a gap of zero means selective deployment recovers within-city quality. We compute the gap on three pixel populations: the foreground class (GT-shadow, confidence = ), the background class (GT-background, confidence = ), and all pixels (confidence = ). All numbers in this subsection come from the 63 LOCO cells (architecture × method × holdout city, for high-resolution); paired image-level bootstrap () gives 95% confidence intervals.
The gap concentrates on the foreground class. On shadow pixels the gap averages +0.136 across the 63 cells. On background pixels it averages +0.006. The asymmetry is consistent across cells: the foreground gap exceeded the background gap in 55 of the 63 cells (sign-test ). Two further measurements describe how the model’s confidence shifts. At the logit level, true-positive logits drop on average by 0.85 nats (the model becomes much less confident on its foreground predictions), while logits on false-positive predictions barely move (−0.15 nats); the drop is larger on the true-positive side in 55 of 63 cells. Calibration error rises asymmetrically in the same direction (Appendix G.6). The LOCO failure is therefore not uniform confidence flattening: the model selectively loses commitment on the class it is trying to detect, while leaving its background commitment intact.
Cross-task scope test on INRIA building footprints. A natural objection is that these patterns are properties of shadow appearance, not geographic transfer. We replicate the protocol on INRIA building extraction [33], a dense-prediction task whose foreground is shape-defined rather than illumination-defined (MAMNet, holdout cities Austin/Chicago/Vienna, matched training recipe). The foreground SP gap and the foreground-vs-background asymmetry both transfer in direction across tasks (sign test at the strongest result the sample size permits, each; Cohen’s ), confirming the SP-gap signature is a property of LOCO across tasks rather than of shadows. Figure 3(a) inset shows the gap distributions side by side.

Figure 3. Diagnostic framework. (a) Risk-coverage curves on foreground pixels (MAMNet, Phoenix held out): LOCO accumulates over the upper bound; FDA narrows it. Bands: 95% bootstrap CIs. Inset: foreground gap replicates on INRIA building footprints under matched LOCO (C/M/P: Chicago/Miami/Phoenix; A/C/V: Austin/Chicago/Vienna). (b) Per-method AURC gap across 9 cells (3 architectures 3 cities): the gap is class-asymmetric, concentrated on foreground (coral) with a pooled background gap (black) 24 smaller over the 63 cells of the 7 methods. (c) Per-image recovery ratio under decoder (Exp A) vs. encoder (Exp B) retraining: the DINOv3OGLANet sign reversal shows the failure locus is architecture-dependent. LOCO: leave-one-city-out; AURC: area under risk-coverage.
A recoverable threshold exists. The selective-prediction framing is only useful if there is some confidence threshold at which LOCO predictions match within-city quality. We find one in 59 of 63 cells; the four exceptions are all on Phoenix, the most distant holdout, under the most aggressive adaptation methods. This makes coverage thresholding, rather than post-hoc temperature scaling, the natural failure-free recovery recipe under geographic shift, and Section 5 builds on it directly. Method-ranking stability across architectures, the post-hoc temperature-scaling alternative, and per-cell threshold values are in Appendix G.6.
Implication. The SP gap is large, foreground-concentrated, hidden by 0/1 evaluation, recoverable by coverage thresholding, and consistent across architectures and across a second dense-prediction task. Section 4.4 asks where in the architecture the failure that produces this gap actually lives.
Architecture-dependent failure locus
Sections 4.2 and 4.3 established that the failure is in feature space and that confidence collapses asymmetrically. Neither result tells us which part of a given model produces the collapse: the encoder, the decoder, or both. This matters because uniform feature-level corrections (ISW whitens, MRFP randomizes) act on a fixed locus, and Table 2 already showed they fail catastrophically on some architectures while helping others. If different architectures carry the failure in different components, that asymmetry is the explanation.
Test design. We localise the failure by targeted retraining. Experiment A freezes the LOCO-trained encoder, reinitializes the decoder, and retrains only the decoder on 25% of the holdout-city training set. Experiment B does the inverse: freezes the decoder, resets the encoder to its pretrained initialization, and retrains only the encoder on the same subset. If a component carries the failure, retraining it on a small target sample should recover most of the within-city gap; retraining the other component should not. Recovery is measured by the per-image ratio defined earlier (training details in Appendix § 14.8).
Mirror-image recovery pattern. Decoder retraining (Experiment A) recovers DINOv3 () and MAMNet () but barely moves OGLANet (, [0.05, 0.21] - CI excludes meaningful recovery). Encoder retraining (Experiment B) shows the opposite pattern: it makes DINOv3 worse than LOCO (), while recovering both CNNs ( MAMNet, OGLANet). All CIs in Appendix § 14.8. The DINOv3-vs-OGLANet sign reversal across the two experiments shows decoder works for one, encoder works for the other, and each architecture is harmed by the other’s fix. This is the central evidence that the failure component is architecture-dependent. Figure 3(c) shows the pattern.
Frozen encoder. If the gap came only from fine-tuning the encoder on two source cities, freezing the encoder would remove it. With the DINOv3 encoder frozen and only the decoder trained, the upper bound still exceeds LOCO by 7.10, 2.67 and 2.75 mIoU on Chicago, Miami and Phoenix (each significant after Holm correction over the three cities; Appendix § 15.§ 3). Both models share the frozen encoder, so this gap arises in the decoder, consistent with the locus Experiment A measures for DINOv3.
Pretraining and fine-tuning regime. DINOv3 also differs from the CNNs in self-supervised pretraining and a low fine-tuning learning rate. We test both on MAMNet with a ResNet-50 encoder by training four versions: ImageNet-supervised or self-supervised DINO pretraining, each fine-tuned at learning rate or (three held-out cities, three seeds each). Encoder retraining recovers all four versions (median 0.80 to 0.87, 95% CIs within [0.76, 0.90]), against −0.13 for DINOv3. The self-supervised, low-learning-rate version recovers less than the supervised, high-learning-rate one (mean per-image difference −0.13, 95% CI [−0.22, −0.05], Holm ), and its recoverable component stays in the encoder. Pretraining scale remains confounded: DINO-ResNet-50 is pretrained on ImageNet-1k (1.28M images) and DINOv3 on LVD-1689M (Appendix § 15.§ 4).
Implication. The recoverable component is architecture-dependent: decoder for DINOv3, encoder for OGLANet, both for MAMNet. This directly explains the catastrophic failures in Table 2. MRFP⁺ perturbs encoder features uniformly without targeting OGLANet’s encoder-locus failure, so it costs OGLANet −14.0 mIoU. ISW whitens features in a way that corrupts the representation DINOv3 depends on, so it costs DINOv3 −18.5 mIoU.
External benchmark: S-EO
We repeat the benchmark and the three probes on S-EO [4]: WorldView-3 imagery of Omaha, Jacksonville and San Diego, four architectures, three seeds, and 3,500 training patches per side (Appendix § 16). Its masks are derived from digital surface models; both sides of every gap share them, so label noise largely cancels in the gap. The upper bound exceeds LOCO in all 12 (architecture, held-out city) pairs (Holm ), by 2.10 to 7.95 mIoU and 5.41 [5.07, 5.75] pooled. Histogram matching lowers mIoU in all 36 vanilla cells, each significant after Holm correction, and the foreground AURC gap exceeds the background gap in all 12 pairs, significantly in 8. Decoder retraining raises foreground IoU in all 18 MAMNet and DINOv3 cells and lowers it in all 9 OGLANet cells; the size of this change tracks the gap of the LOCO model whose decoder is retrained (), while its sign separates by architecture even between DINOv3 and OGLANet models with similar gaps. The background probability also shifts under LOCO in 10 of 12 pairs, so on S-EO the confidence change is not confined to the shadow class. The gap shows no dependence on the sun-elevation distance between test patches and training scenes (cluster-robust , three cities).

Figure 4. SIB module and gap to the upper bound. (a) Spectral Information Bottleneck (SIB) plug-in module. Three color-tinted zones group components by the diagnostic finding they target: architecture-aware preprocessing (FDA on CNN inputs, SAG on U-Net skips, both bypassed for ViT); spectral bottleneck on the encoder output (Haar split of features into content and boundary subbands, content compressed strongly, boundaries gently, passed through); deployment-time coverage threshold . Component-to-finding mapping in § 5.1. (b) Per-method gap to the within-city upper bound across 9 leave-one-city-out (LOCO) cells. Open markers: mean gap with 95% CI; filled: worst cell; dashed grey line: Vanilla mean gap. Values in Table 44. FDA: Fourier Domain Adaptation; SAG: Skip Attention Gates.
Case Study: Spectral Information Bottleneck (SIB)
Section 4 characterized three failure modes under geographic transfer. We now ask whether this characterization translates into design guidance. We present SIB, a plug-in module that maps each finding to a specific design choice. SIB has the smallest mean gap to the within-city upper bound of the ten methods we evaluate, level with FDA. SIB is designed from the same LOCO cells on which we evaluate it, so its results show that the diagnosis yields a usable design and do not test the diagnosis independently.
Module design
Feature-space failure (§ 4.2) → band-asymmetric compression. Based on 4.2, the intervention must operate on feature space. We need to compress domain-specific content while preserving boundary detail. Haar wavelet decomposition gives this separation directly: the low-frequency subband carries surface appearance and city-specific texture; the high-frequency subbands , , carry shadow geometry. Recent VFM-based domain generalization work confirms this content/style separation in feature space [26]. We apply a variational information bottleneck [34] to aggressively and to , gently. passes through. Aggressive compression on removes city-specific content, while gentle compression on the boundary subbands prevents the bottleneck from erasing the edge structure shadow detection requires.
Architecture-dependent locus (§ 4.4) → skip gates and CNN-only preprocessing. The bottleneck acts on encoder output. In U-Net, skip connections route encoder features directly to the decoder, bypassing the bottleneck. We add Skip Attention Gates: channel-wise sigmoid gates on each skip, identity-initialized so they default to a no-op. The model learns to suppress city-specific channels routed around the bottleneck. CNN encoders also absorb pre-encoder pixel statistics no feature-level intervention can reach. We add Fourier Domain Adaptation [20] as input preprocessing on the CNN branch only. ViT encoders carry their recoverable component in the decoder (§ 4.4) and receive neither component.
Class-asymmetric SP failure (§ 4.3) → coverage thresholding. The model’s confidence ranking survives geographic shift even when calibration fails. We exploit the surviving ranking with a deployment-time recipe: per cell, fit as the largest (coverage) at which selective shadow error on the source-city validation split is at most , with fixed globally, i.e., the criterion is to halve the source-val full-coverage error. Apply at deployment on the held-out city. The procedure uses no held-out labels. Rank preservation guarantees per-image selective error is bounded by full-coverage error, so the population test measures the magnitude of the recovery rather than its direction. Training loss and full hyperparameters in Appendix § 17.
Results
Gap to the upper bound. Over the 9 cells (3 architectures × 3 held-out cities; Table 4), SIB has the smallest mean gap to the within-city upper bound of Vanilla and the eight adaptation methods, 6.42 mIoU (95% CI [5.01, 7.96]), level with FDA at 6.47 (difference +0.05, [−0.42, +0.49]). Its worst-case gap, 10.41 mIoU (OGLANet, Miami), is second to FDA’s 9.10 and cannot be separated from it (difference +1.31, [−1.01, +3.65]). FDA and SegDeSiC use unlabeled target-city images in every cell, as SIB does on its CNN backbones through FDA. SIB leaves a significant gap in all 9 cells, and on S-EO its worst cell scores 1.57 mIoU below Vanilla’s (Appendix § 16). The deployment-time coverage recipe reduces selective shadow error on all 9 cells (mean , paired Wilcoxon ), stable across (Appendix § 20).
| Status (color) | Trigger | Kept segment | Annotator action | Share |
| Correct (green) | , IoU | Verify, usually skip | 65% | |
| Potential_Annotated (yellow) | , IoU | (edit shape) | Adjust boundary | 21% |
| Potential_Model (blue) | has no overlapping | if validated | Confirm or reject; assign type | 9% |
| Incorrect (red) | has no overlapping | if validated | Verify segment is shadow | 5% |
Table 4. Phase 3 adjudication rules. Each (annotation, prediction) pair on an image is assigned one of four statuses based on their IoU and overlap, and routed to the corresponding annotator action. : human annotation; : model prediction. Distribution column reports the share of all Phase 3 segments falling into each status.
Component ablations. Table 3 reports the change in mIoU when each SIB component is removed, against each architecture’s full-model reference from the same training batch, with Holm correction over the 28 ablation tests. The largest significant changes are on OGLANet, where removing SAG costs 2.06 mIoU, compressing in place of costs 1.99, and removing the content VIB costs 1.01. On MAMNet the only significant row removes FDA preprocessing together with the contrast stage (−4.42 mIoU), and removing SAG changes mIoU by +0.23 (not significant). On DINOv3, removing the Haar decomposition gives −23.49 mIoU, but that run predicts no shadow on nearly every Chicago and Miami image, so the row records a training collapse and does not isolate the decomposition’s contribution; the content-VIB and inverse-subband rows on DINOv3 are not significant after correction. Appendix § 19 lists every row together with the training differences between each ablation and its reference.
| Variant | MAM | OGL | DIN |
| − SAG | +0.23 | −2.06** | N/A |
| − FDA, contrast | −4.42** | −0.87 | N/A |
| − Haar | +0.11 | −0.63 | −23.49***‡ |
| − VIB on | +0.29† | −1.01** | −0.71 |
| VIB on (inv.) | −0.29 | −1.99** | −0.74 |
Table 3. SIB component ablations. Mean mIoU when each component is removed (or, for the inverse-evidence row, replaced), against each architecture’s full-model reference from the same training batch, pooled over the three held-out cities ( paired images). Stars: Holm-adjusted over the 28 ablation tests, * , ** (paired image bootstrap, ). †MAMNet row removes every VIB. ‡Run predicts almost no shadow. Full table in Appendix § 19.
Discussion and Conclusion
We introduced ShadowTransfer, a 4,500-image human-verified benchmark for geographic transfer in overhead shadow detection across three U.S. cities and two native resolutions, with a leave-one-city-out protocol and paired bootstrap reporting. Its three-probe framework indicates that the failure operates in feature space (§ 4.2), appears as a class-asymmetric collapse of foreground confidence that 0/1 metrics under-report by 2.3× (§ 4.3), and is recoverable through an architecture-dependent component: decoder for DINOv3, encoder for OGLANet, both for MAMNet (§ 4.4). Four of five selective-prediction findings replicate on INRIA building footprints. The gap remains significant in every 900-image cell and on the external S-EO benchmark. As a case study, SIB has the smallest mean gap of the ten methods we evaluate (6.42 mIoU), level with FDA.
The benchmark covers three U.S. cities, RGB-only NAIP imagery, a single fall season, and four architectures. Non-grid morphologies, multispectral sensors, and untested architectures (e.g., SegFormer) remain open. We characterize where the failure occurs, how it distorts predictions, and which component recovers it, but do not isolate a causal mechanism that connects these characterizations. DINOv3’s decoder locus also cannot be attributed to its ViT architecture alone, since no CNN in our tests matches its pretraining scale. We release ShadowTransfer, the LOCO splits, training code, and trained models as an evaluation instrument to be read with the three probes: its LOCO gaps measure geographic transfer risk in overhead dense prediction, and its two native resolutions support diagnosis of cross-resolution transfer failure. Together, the benchmark and framework give the field a shared instrument for the question that matters at deployment: not just whether a model works, but where and on which class it will fail when the city changes.
Acknowledgments and Disclosure of Funding
This work received no external funding. This work used Delta at the National Center for Supercomputing Applications (NCSA) through allocation CIS260282 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program [35], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. This research used the Delta advanced computing and data resource, which is supported by the National Science Foundation (award OAC 2005572) and the State of Illinois. Delta is a joint effort of the University of Illinois Urbana-Champaign and its National Center for Supercomputing Applications. The authors declare no competing interests.
Datasheet
Following [36], this datasheet documents SHADOWTRANSFER. Items already detailed in the main paper or other appendices are cross-referenced rather than repeated.
Motivation
Purpose. SHADOWTRANSFER was created to enable measurement of geographic transfer in overhead shadow detection. No existing overhead shadow dataset combines human-verified labels, multiple cities, native multi-resolution coverage, and a defined cross-city evaluation protocol (Table 1); the most recent overhead-shadow survey [19] identifies cross-dataset generalization as an open problem.
Creators and funding. Shagun Mittal, Raymond A. Yeh and Satish Ukkusuri (Purdue University). The work received no external funding; compute was provided through ACCESS allocation CIS260282.
Composition
Instances. 4,500 image–mask pairs: 3 cities (Chicago, Miami, Phoenix) × 2 native NAIP resolutions (0.3 m/px, 0.6 m/px) × 750 patches. Each instance is a RGB orthoimage patch and a co-registered binary shadow mask; six-category multiclass masks accompany the patches where reliable typing was possible (App. § 10.§ 1).
Sample vs. full. A spatial sample of NAIP imagery within each city’s administrative boundary; protocol in App. § 10.§ 3.
Files per instance. 8-bit RGB .png; 8-bit binary mask .png ({0, 255}); optional 8-bit multiclass mask (IDs 0–6); and a per-patch metadata record (city, native resolution, lat/lon, source NAIP tile, paired-coordinate id, shadow types present).
Labels. Pixel-level binary shadow (primary); six typed shadow categories (secondary, App. § 10.§ 1). Sparse types (vehicle-cast, thin-linear) are not recommended as primary evaluation targets.
Relationships between instances. 300 paired patches per city share a pair_id linking the two native resolutions at identical ground coordinates (App. § 10.§ 3).
Recommended splits. Per-city: 450 train / 150 val / 150 test, fixed at the benchmark level. Cross-city: the LOCO protocol (§ 3.3) defines three rotations; pre-built folds are released alongside the per-city raw data.
Errors and noise. Shadow boundaries are inherently imprecise; we recommend boundary-tolerant mIoU (±2 px, App. § 12). A ≈ 3% residual inter-annotator disagreement persists after three-phase QC (App. § 9).
Self-containment. Self-contained; underlying NAIP imagery is U.S. public domain (§A.6).
Sensitive content or identifiable individuals. None. Patches are land-cover orthoimagery; no PII, no identifiable individuals, no confidential information.
Collection Process
Acquisition. Source NAIP tiles obtained from USDA-FSA APFO; patches cropped within each city’s administrative boundary under one-patch-per-tile sampling to limit spatial autocorrelation (App. § 10.§ 3).
Annotation labor. Seven trained annotators produced masks across 45 sessions in CVAT [37] with SAM-assisted boundary proposals [38]; annotators were paid INR 20 per accepted image, and a second annotator who reviewed an image was paid a further INR 20, about $0.50 per image in total (Appendix § 9.§ 3).
Time frame. A single fall NAIP acquisition window per city; annotation completed across the 45 sessions.
IRB / ethics review. Not required: no human subjects, no PII.
Preprocessing, Cleaning, Labeling
Preprocessing. Patches cropped to at native resolution; no radiometric, color, or geometric transforms applied. Source tile IDs and patch coordinates are released, so raw inputs can be re-derived from NAIP.
Labeling. Three-phase pipeline (App. § 9): SAM-assisted manual draw, model-in-the-loop acceleration, and a retroactive QC pass that intersects each segment with a final-model prediction. Frozen gold masks (5 per 100-image batch) drive Cohen’s monitoring; phase-3 mean .
Uses
Primary use. Benchmarking shadow detection on overhead aerial imagery, with explicit measurement of geographic transfer (LOCO) and in-domain performance per city.
Already used for. Baseline, diagnostic, and case-study results in this paper, including cross-task replication on INRIA building footprints (§ 4.3, App. § 14.7).
Other suggested uses. Building footprint and façade extraction (occlusion priors), shadow-removal research, domain generalization for dense prediction, and urban-morphology or solar-geometry analysis from the released coordinates and shadow-type labels.
Uses not supported. Generalization beyond grid-pattern North American urbanism, beyond NAIP RGB sensor characteristics, or across seasons (the dataset captures a single fall window) is not supported. The 0.3 m/px and 0.6 m/px subsets are separate native acquisitions; users studying resolution transfer should not synthesize one from the other.
Distribution
Where. Hugging Face dataset record at https://huggingface.co/datasets/shagunmittal/ShadowTransfer; DOI 10.57967/hf/11483 [39]. Code at https://github.com/shagun8/shadowtransfer.
When. Public release at publication; the version corresponding to paper results is Hugging Face revision 5bef90e, tagged v1.0.
License. Annotations and metadata are released under CC BY 4.0. NAIP imagery is U.S. public domain (USDA-FSA APFO attribution requested for derived products). Code is released under the MIT license.
Export or regulatory restrictions. None known.
Croissant metadata. A machine-readable MLCommons Croissant description, including the Responsible AI fields, is released with the dataset at https://huggingface.co/datasets/shagunmittal/ShadowTransfer/resolve/main/croissant.json and is consumable by mlcroissant, TFDS, and any Croissant-aware loader.
Maintenance
Maintainer. Shagun Mittal ([email protected]); correspondence is also handled via the Hugging Face dataset record (issues/discussions) and the GitHub repository.
Versioning. Tagged releases on Hugging Face; v1.0 freezes the splits used for paper results. Errata are posted as new tags; older tags remain accessible.
Contributions. External contributions accepted via pull request on the GitHub repository, tagged separately from v1.0.
Retention or individual rights. N/A: no individuals or personal data are present.
Annotation Pipeline Details
This appendix documents the three-phase pipeline summarized in § 3.2: the per-phase tooling (Figure 5), the Phase 3 adjudication rules that reconcile human labels with model predictions (Table 4), and the gold-image protocol and Cohen’s trajectory that monitor annotator drift across batches (Table 5). Together these mechanisms support the human-verified label claim made against S-EO [4] in Table 1.

Figure 5. Three-phase annotation pipeline. Phase 1 (SAM-assisted manual labeling) trains an initial model; Phase 2 (model-in-the-loop) accelerates throughput ; Phase 3 (retroactive quality control) reconciles labels against final model predictions. Each phase’s output seeds the next, jointly solving scale and label quality. Inter-annotator agreement statistics and adjudication rules are in Appendix § 9.
| Phase | Sessions | Mean | Range across annotators | First-submission accept |
| Phase 1 (manual) | 1–10 | 0.71 | [0.62, 0.78] | 84% |
| Phase 2 (model-in-the-loop) | 11–45 | 0.84 | [0.79, 0.89] | 96% |
| Phase 3 (post-adjudication) | — | 0.89 | [0.85, 0.92] | — |
Table 5. Inter-annotator agreement across phases. Mean Cohen’s on gold images across the seven annotators, with the across-annotator range and the per-batch first-submission acceptance rate (§B.2). Phase 3 reports post-adjudication agreement against the final reconciled corpus.
Three-Phase Pipeline
Seven trained annotators produced shadow masks in CVAT across 45 sessions of 105 images each (100 working images + 5 gold images with frozen reference masks; §B.2). Sessions proceed in three sequential phases (Figure 5); each phase’s output seeds the next.
Phase 1: SAM-assisted manual labeling. Annotators draw shadow masks on raw 384×384 patches. The CVAT AI Interactor [37], powered by the Segment Anything Model (SAM) [38], proposes a candidate boundary from a single annotator click; the annotator accepts, refines, or rejects it. Manual polygon drawing is available for thin-linear and irregular shadows where SAM under-segments. Each segment receives a shadow-type label from the six-category taxonomy (§C.1). Phase 1 spans the first 10 sessions and is restricted to Chicago and Miami imagery balanced across the two native resolutions; Phoenix is admitted only after the label distribution stabilizes (gold-image for a majority of annotators). Mean throughput is 4.8 minutes per image after a one-session per-annotator ramp-up.
Phase 2: Model-in-the-loop acceleration. After Phase 1, all labels train an interim shadow detector. The detector is one of {MAMNet, OGLANet, DINOv3}, drawn at random for each new 100-image batch; rotating architectures limits annotator entrainment to any single model’s failure modes. Predictions from the interim model are loaded into CVAT as the starting layer; annotators delete false positives, refine boundaries, and add missed segments rather than redraw from scratch. The interim model is retrained at session boundaries on the cumulative labeled corpus. Throughput rises to 75 seconds per image, a acceleration over Phase 1.
Phase 3: Retroactive quality control. After all working sessions complete, a final shadow detector is trained on the full Phase 1+2 corpus and used to revisit every annotated segment. For each image we compute the per-segment Intersection-over-Union (IoU) between each human annotation and each model prediction . Each (annotation, prediction) pair receives one of four statuses (Table 4). Statuses are color-coded in CVAT so annotators triage at a glance: green segments are skipped; yellow, blue, and red segments are routed to a second annotator pass for adjudication. The status taxonomy is released with the dataset for downstream provenance analyses; for the LOCO benchmark (§ 3.3), all final segments are merged into the binary foreground class regardless of their adjudication route.
Gold-Set Protocol and Inter-Annotator Agreement
Gold images. Every 100-image working batch is paired with 5 gold images that carry frozen reference masks. Gold images are drawn from previously closed batches that have already cleared Phase 3, and their reference masks are independently re-adjudicated by two senior annotators before the gold image enters circulation. Once frozen, a gold mask is not modified. Gold images are interleaved with working images and are not visually distinguishable from them in CVAT, so annotator behavior on gold matches their behavior on the rest of the batch.
Cohen’s trajectory. For each annotator and each batch, we compute Cohen’s on binary shadow/background agreement between the annotator’s mask and the frozen gold reference, pooled over the 5 gold images. Mean across the seven annotators rises from Phase 1 to Phase 2 as model seeding stabilizes the label distribution, and rises further after Phase 3 adjudication (Table 5). All Phase 2 and Phase 3 values fall within the substantial-to-almost-perfect range () [40].
Acceptance gates. A batch is accepted when its mean gold mask-IoU against the frozen reference is at least 0.90; spot-checked on 5–10 randomly selected working images by a senior annotator before entering the corpus. A batch with mean gold IoU between 0.80 and 0.90 is conditionally accepted: the annotator receives written feedback on the failed gold images, and the next batch dispatched to that annotator includes 5 additional gold images. A batch with mean gold IoU below 0.80 is rejected and the failed images returned for redo; a second failed submission triggers a paired-review session with a senior annotator before the next solo batch is dispatched. Across all sessions, 96% of batches were accepted on first submission, and 4% were resubmitted.
Annotator Recruitment and Instructions
Annotation was paid, on-demand work carried out in India. The full text of the recruitment and task instructions given to annotators follows; amounts are in Indian rupees (INR). In practice every accepted image was paid INR 20, and each review by a second annotator was paid a further INR 20.
Mark Shadow Areas on Images (Web Tool)
Why this work matters. We’re building a high-quality research dataset that helps improve mapping and city planning. Your careful work directly contributes to real-world research used by universities and labs.
What you’ll do. Open a small set of images (satellite view) in a web tool. Mark areas that are in shadow by drawing/adjusting shapes. Save and move to the next image. That’s it. No coding. No prior mapping experience required.
Training & support. Short written guide with pictures. Quick sample demo. We provide fast feedback so you can improve and earn the quality bonus.
Payment (per accepted image). INR 20 per image. Each session is ∼100 images, so the typical total is INR 1,500–2,000. Payments are made after review of each session.
How we review quality. Each batch passed through a validation AI model. We do a short spot-check on a few others. If any images need fixes, we’ll send them back with quick notes; you’ll be paid for the images once approved.
Selection process (short test). You’ll get a mini test (5 images) in the same web tool. If the test looks good, we’ll invite you for paid work immediately. Top performers get more sessions each week.
Per-City Dataset Statistics and Shadow Taxonomy
This appendix documents the six-category shadow taxonomy referenced in § 3.2 (§C.1), the per-city and per-resolution distributional statistics that motivate the LOCO design (§C.2), and the spatial sampling protocol used to draw patches within each city (§C.3).
Shadow Taxonomy
Each annotated shadow segment carries one of six type labels. The taxonomy is released with the dataset for downstream analyses; the LOCO benchmark (§ 3.3) collapses these to a binary mask. Definitions and inclusion rules are in Table 6. Annotators applied a single inclusion rule across all six types: a region is shadow if it is solar-directional and consistent with local scene geometry. Sensor artifacts, dark water, and cloud shadows lacking geometric consistency are excluded.
| Category | Definition and inclusion rule |
| Building / canyon | Cast shadows from buildings and built structures, including narrow inter-building canyons. |
| Under-structure | Persistent occlusion shadows beneath elevated structures (bridges, over-passes, awnings, carports). |
| Tree-canopy dapple | Shadow patterns cast by tree foliage, including dappled shadow on the ground beneath the canopy. |
| Topography-cast | Shadows produced by terrain elevation changes (hills, embankments, cuttings). |
| Vehicle-cast | Shadows cast by cars, trucks, and other ground vehicles. |
| Thin linear | Shadows from thin vertical structures (poles, fences, antennas, wires) whose width is on the order of a few pixels. |
Table 6. Six-category shadow taxonomy. Every annotated segment receives exactly one label. The binary LOCO benchmark merges all six into the foreground class.
Per-City and Per-Resolution Statistics
Figure 6 reports the distributional properties of the benchmark. Four observations bear on the LOCO design and on results reported in § 4.

Figure 6. SHADOWTRANSFER dataset statistics. (A) Sample patches across city-resolution pairs with shadow annotation overlays. (B) Shadow-coverage distribution per city (high-res); Phoenix patches are predominantly high-coverage ( shadow area), driving the surface-albedo shortcut. (C) Median shadow blob size per category and city, showing strong city-level variation in shadow morphology. (D) Shadow-category distribution per city. (E) Mean grayscale intensity distributions per city (-, -); Phoenix’s lower standard deviation reflects more uniform high-albedo surfaces.
Shadow coverage is city-dependent (panel B). Phoenix is dominated by low-coverage patches: 49% of high-resolution images contain 5–15% shadow area and only 9% exceed 30%. Miami and Chicago are shifted toward higher coverage (50% and 31% of images above 30% coverage, respectively). Per-image annotation density follows the same trend (29.7 blobs/image in Phoenix, 48.4 in Miami, 38.6 in Chicago).
Shadow morphology varies by city (panels C, D). Chicago has the largest median tree-canopy blobs (255 px) and under-structure blobs (193 px), reflecting a denser urban grid with mature street trees. Phoenix shows the smallest under-structure blobs (72 px) and a vehicle-shadow share of 7%, the highest among the three cities. Miami’s distribution is dominated by tree-canopy shadow (47% of segments) consistent with its dense subtropical canopy. Chicago is dominated by building shadow (53%).
Surface intensity is city-dependent (panel E). Mean grayscale intensity distributions overlap in mean () but differ in spread: Phoenix () is roughly half as wide as Miami (). The narrow Phoenix distribution is consistent with uniform high-albedo desert surface; the wide Miami distribution reflects a mix of bright sand, asphalt, and dark vegetation.
Native resolution affects blob count, not city ranking (panel A). The same ground footprint contains substantially more shadow blobs at than at (+85 to +100 blobs in the illustrated examples) because the lower-resolution patch covers more area on the ground. The relative ordering across cities is preserved across resolutions.
These statistics establish that the three cities differ along orthogonal axes (coverage, morphology, surface intensity) rather than along a single appearance axis, supporting the LOCO claim in § 3.1 that the protocol measures transfer across genuine appearance shifts.
Spatial Sampling Protocol
Patches are drawn from NAIP orthoimagery in two streams: random patches that provide diverse independent samples, and paired patches that share ground coordinates across the two native resolutions.
Random patches. Within each (city, resolution) pair, target patch counts are distributed across NAIP tiles overlapping the city administrative boundary. If the target count is at most the tile count, tiles are sampled without replacement and each sampled tile contributes one patch; otherwise the count is distributed evenly with the remainder allocated to randomly chosen tiles. Within each selected tile, a candidate center is sampled uniformly inside the tile geometry and accepted only if (i) the point lies within the city boundary and (ii) a window centered at the point fits entirely inside the tile (a 192-pixel buffer to all edges). Up to 50 candidates are tried per slot; unfilled slots are skipped. The one-patch-per-tile rule limits spatial autocorrelation between samples.
Paired patches. For each city, 150 ground coordinates are drawn under the random-patch procedure on the mid-resolution stream ( m/px), with the additional constraint that each new center is at least 384 pixels (one patch width) from every previously accepted center in the same tile. The accepted longitude/latitude coordinates are written to a shared registry and reused for the high-resolution stream ( m/px): for each saved coordinate, the high-resolution NAIP tile containing the location is identified and a window is extracted centered on the exact same geographic point. Each paired patch carries a shared pair_id linking its two resolutions. Paired patches support resolution-matched analyses; results reported in the main paper use the combined random+paired pool unless stated otherwise.
Reproducibility. Sampling is organized into sequential sessions (40 random and 10 paired patches per resolution per session), each with a distinct random seed. Per-patch metadata (city, scale_gsd_m, lon, lat, source_tile_id, pair_id) is released with the benchmark, enabling exact patch-level reproduction of the train/val/test splits.
Split Distribution Characterization
Overhead remote-sensing benchmarks rarely verify that train, validation, and test splits are distributionally matched within each domain, despite documented evidence that within-domain spatial autocorrelation can inflate apparent performance by up to 28% [41]. We close this gap for SHADOWTRANSFER with two complementary diagnostic suites: omnibus feature-space tests that ask whether the splits are distinguishable at all (§D.1), and a task-relevant feature panel that localizes any divergence to interpretable axes (§D.2). The diagnostics confirm that the LOCO Upper Bound measures within-city generalization rather than within-city heterogeneity, and they isolate one bounded caveat (§D.3) for transparent reporting.
Omnibus Diagnostics
For each of the six city, resolution pairs we run four within-city tests comparing train against test. (D1) Two-sample Kolmogorov-Smirnov on label statistics (per-image shadow coverage, blob count, mean blob area), reporting KS distance with image-level bootstrap CI (); at flagged. (D2) Domain-classifier covariate-shift test: ImageNet ResNet-50 features feed a logistic regression trained to distinguish train from test images of the same city under 5-fold cross-validation; ROC-AUC flagged. (D3) Bootstrap maximum mean discrepancy (MMD) with RBF kernel under the median-heuristic bandwidth on DINOv2 features ( per split, ); flagged. (D4) Spatial coverage: per-split longitude/latitude range, mean test-to-nearest-train distance, and KS on each coordinate.
| City | Res. | Mean | AUC | MMD | Spat. |
| Chicago | High (0.3 m) | 0.18† | 0.73† | 0.02† | 0.21† |
| Chicago | Mid (0.6 m) | 0.09 | 0.56 | 0.41 | 0.08 |
| Miami | High (0.3 m) | 0.07 | 0.54 | 0.62 | 0.11 |
| Miami | Mid (0.6 m) | 0.08 | 0.58 | 0.34 | 0.13 |
| Phoenix | High (0.3 m) | 0.08 | 0.51 | 0.71 | 0.07 |
| Phoenix | Mid (0.6 m) | 0.13 | 0.62 | 0.18 | 0.10 |
Table 7. Within-city split diagnostics. Mean D: mean KS distance over the three label statistics (D1). AUC: domain-classifier ROC-AUC (D2). MMD p: bootstrap -value (D3). Spat.: KS on (lon, lat) (D4). †: flagged (, AUC , or ). Five of six cells show no significant mismatch; Chicago highres is the single exception.
Phoenix shows the cleanest splits of any city by every diagnostic. The single flagged cell, Chicago highres, traces to a structural property of the sampling protocol: test images come from the paired-coordinate subset (§C.3), which can only be drawn from regions where NAIP provides overlapping coverage at both native resolutions. In Chicago that dual-coverage zone is geographically narrower than the random-train footprint (test latitudes start at 41.687 vs. 41.748 for train), producing the spatial concentration captured by D4 and propagating to the feature-space tests.
Per-Feature Divergence Panel
The omnibus suite answers whether train and test are distinguishable; for practitioners the more useful question is which task-relevant features are mismatched and by how much. We compute the train-vs-test KS distance separately on a panel of nine features that mechanically affect mIoU: shadow coverage fraction, blob count per image, mean blob area, small-object fraction, shadow-category mix (total variation in place of KS), and per-image mean intensity for each of R/G/B and the grayscale standard deviation.
The panel reproduces the omnibus picture at finer resolution and quantifies the size of the divergences. Chicago highres flags eight of nine features: shadow coverage shifts from 0.103 in train to 0.123 in test (+19% relative), median blob area shifts by +42%, and blob count by +18%. These shifts are consistent with the dual-coverage geographic concentration documented above, the paired test footprint over-samples denser-shadow neighborhoods in central Chicago. Phoenix midres flags five features, but each at the threshold (, relative shifts 2–5%), an order of magnitude smaller than the Chicago highres effect. Phoenix highres flags zero features and posts the lowest mean in the benchmark.
| City | Res. | Max | Mean | Sig./9 |
| Chicago | High | 0.21 | 0.14 | 8 |
| Chicago | Mid | 0.12 | 0.07 | 0 |
| Miami | High | 0.13 | 0.08 | 1 |
| Miami | Mid | 0.14 | 0.09 | 2 |
| Phoenix | High | 0.12 | 0.08 | 0 |
| Phoenix | Mid | 0.15 | 0.11 | 5 |
Table 8. Per-feature train-vs-test divergence. Max D: largest single-feature KS distance over the nine-feature panel. Mean D: panel average. Sig./9: count of features with and . Chicago highres flags eight of nine features at substantial magnitudes; remaining cells flag at most five features, all at the threshold.
Implications for LOCO Reporting
Two conclusions follow.
Phoenix’s LOCO degradation reflects cross-city transfer, not split heterogeneity. Phoenix’s within-city splits are statistically matched on every diagnostic, and the LOCO protocol does not draw on Phoenix’s training split when Phoenix is held out. Phoenix’s lower Upper Bound is partly mechanical: shadow coverage in Phoenix (5.7–6.7%) is roughly half that in Chicago/Miami (8.7–15.1%), the small-object fraction (0.80–0.82 vs. 0.41–0.50) is nearly twice as high, and mean R-channel intensity differs by units (§C.2). All three properties suppress mIoU regardless of model quality. The cross-city gap reported in Table 2 is therefore a genuine transfer phenomenon, not a split artifact.
Chicago highres carries a known, bounded caveat. The paired-coordinate test set covers a narrower geographic band than the random training set, producing the +19% shift in test-set shadow coverage documented above. This affects Chicago’s Upper Bound (and therefore the magnitude of in cells where Chicago is held out) but does not introduce circularity into the held-out-city evaluation, since LOCO never co-locates Chicago train and Chicago test. The direction of bias is conservative for our headline claim: Chicago already posts the highest Upper Bound in Table 2, so a slight under-estimate of its true Upper Bound would only widen the reported LOCO gap. We disclose the caveat explicitly and release per-patch metadata (pair_id, lon, lat, source_tile_id) so downstream users can re-split if a fully matched protocol is required for their analysis.
Metric Definition
This appendix formalizes the boundary-tolerant mIoU used as the primary metric in § 3.3: the -pixel don’t-care band, the per-image IoU on the valid region, the aggregation across images, and the rule for degenerate cases.
Don’t-care band. For a binary ground-truth mask and a tolerance pixels, let be an elliptical structuring element of radius (OpenCV MORPH_ELLIPSE, kernel size ). The don’t-care band and the valid evaluation mask are
Pixels inside are “definite shadow”; pixels outside are “definite background”; pixels in are excluded from numerator and denominator alike, so errors inside the band are neither rewarded nor penalized. Setting recovers the strict variant (, covers the whole image).
Boundary-tolerant IoU. For a binary prediction , the per-image foreground IoU on the valid region is
The background IoU is defined in the same way with the two labels exchanged. Throughout the paper, “mIoU” denotes the two-class mean, , computed per image; we report the foreground IoU of Equation (2) separately where a statement concerns the shadow class alone.
Aggregation. We report the per-image mean across the test images per cell,
We do not micro-pool TP/FP/FN counts across images. Per-image macro aggregation gives each patch equal weight regardless of shadow coverage, which matters because coverage varies by city (Appendix § 10.§ 2); under micro-pooling, coverage-heavy patches would dominate the cell mean and conflate geographic transfer with shadow-mass distribution. Per-image scalars are also required by the paired-bootstrap protocol in § 3.3: the bootstrap resamples per-image values to construct cell-level confidence intervals.
Degenerate cases. Two cases require an explicit rule. (i) Components consumed by the band. If a connected shadow component is thin or small enough that removes it entirely, the component contributes zero pixels to both numerator and denominator and is silently dropped from Equation (2). We track such components in the size-stratified “tiny” bin (Appendix § 14); under this affects components narrower than pixels in any direction. (ii) No foreground in . If an image has no shadow on , the IoU denominator is zero; we assign IoU when the prediction also has no positives on (perfect agreement) and IoU otherwise. This case affects 13.0% of test images in SHADOWTRANSFER (117 of 900 at ) and follows the standard convention in binary segmentation.
Rationale for the -pixel tolerance. Annotators draw boundaries on the pixel grid and SAM-proposed boundaries inherit the same quantization, but the true projected shadow edge depends on the sub-pixel position of the occluder and the sun angle, so pixel-exact boundary agreement is unachievable in principle. At 0.3 m/px, two pixels correspond to m on the ground — smaller than the typical width of a tree trunk or roof eave, comparable to the annotator-quantization scale on gold images, and well below the scale of genuine geometric errors that LOCO transfer should expose. Every number in the paper uses .
Training configurations
This appendix documents the architecture and training configuration for each of the four architectures (MAMNet, OGLANet, DINOv3, CrossEarth) and the eight adaptation methods (FDA, SegDeSiC, IIM, ISW, MRFP+, FADA, BSM, SAR). SegDeSiC, IIM, ISW, MRFP+ and FADA are not run on CrossEarth, whose query-based Mask2Former decoder has no per-pixel head these methods are defined on. Each adaptation block lists only those hyperparameters that differ from the corresponding base configuration; values matching the base are omitted.
MAMNet
Architecture. ResNet-34 [42] backbone with pretrained weights ResNet34_Weights.IMAGENET1K_V1 (torchvision); conv1 stride changed from 2 to 1 and maxpool replaced by nn.Identity, so the encoder returns five feature maps at channels with output strides relative to the 384 × 384 input (feat5: 512 × 24 × 24). The MSCAF module operates on feat5 before the decoder. The decoder comprises four DecoderBlock stages (512→256→128→64→64); each block applies Criss-Cross Attention (CCA) to the incoming skip-connection feature, bilinearly upsamples the previous decoder map, concatenates, and refines with two 3 × 3-BN-ReLU convolutions. CCA uses 1 × 1 projections to channels for , and for ; row- and column-wise attention maps are summed and gated by a learnable scalar (init. 0): out = . Three auxiliary branches (AUX1–3) attach to decoder outputs at 256/128/64 channels (Conv3×3-BN-ReLU-Dropout(0.3)-Conv1×1, bilinearly upsampled to input resolution). Input is 384 × 384 four-channel RGBC; the contrast-channel conv1 weight is initialised as the channel-mean of the pretrained RGB weights.
MAMNet + FDA
MAMNet + FDA. FDA [20] is applied at the input stage, before resize and normalisation: each source training image (uint8 PIL numpy) is passed through FDATransform on CPU. Per-channel 2-D FFT separates amplitude and phase; the source amplitude rectangle of half-width centred on the shifted spectrum is replaced element-wise by the corresponding target rectangle (no blending), and the image is reconstructed from the mutated amplitude with the source phase , clipped to . With and , the swapped region is per channel. Validation and test images are unmodified. Run in LOCO mode only; the target pool is up to 100 unlabelled images sampled uniformly per iteration from the held-out city’s train/images.
Table 10: MAMNet + FDA hyperparameters; settings matching Table 9 are omitted.
| Hyperparameter | Value |
| FDA low-freq. ratio | 0.01 |
| Target pool size | images |
| Target domain | Held-out city train/images (unlabelled) |
| Application point | Input (pre-resize, train only) |
| Target sampling | Uniform random per iteration |
| Early-stopping patience | 15 epochs |
| Modes run | LOCO only |
Table 10.
MAMNet + SegDeSiC
MAMNet + SegDeSiC. SegDeSiC [24] attaches a geographic coordinate branch to feat5 (): global average pool MLP (, ReLU + Dropout(0.3) between layers) normalisation, predicting a 40-d GRID encoding of produced by sphere projection (, ) and across 10 log-spaced wavelengths (). The domain loss is , batch-mean. In LOCO mode the loss is computed on both the labelled source batch and an unlabelled target batch (held-out city, no segmentation labels); both terms are summed before the scaling. No coordinates are passed at validation or test time.
| Hyperparameter | Value |
| Geographic branch | |
| Attachment point | feat5 (512×24×24) |
| GRID scales / range | 10 / log-spaced |
| Encoding dim | 40 (4×10) |
| MLP hidden dim / dropout | 256 / 0.3 (both layers) |
| Loss | |
| Domain loss | , batch mean |
| (domain weight) | 0.5 |
| Total loss | |
| CE ignore index | 255 |
| Optimisation | |
| Learning rate | |
| Early-stopping patience | 15 epochs |
| Data | |
| Target domain (LOCO) | Held-out city train/images (unlabelled) |
| Geo. metadata | Per-image lat/lon in mapping.json |
| Modes run | LOCO only |
Table 11. MAMNet + SegDeSiC hyperparameters; settings matching Table 9 are omitted. †No submit script provided; argparse default may have been overridden at launch.
MAMNet + IIM
MAMNet + IIM. IIM [21] is an input-level preprocessor placed before the backbone: input RGB is denormalised, log-transformed, and convolved by learnable kernels in three cross-channel pairs (, , ), producing 24 channels. A FuseConv block fuses these with the original RGB through two parallel paths (each BN-ReLU), concatenation, and two convolutions, returning a 3-channel output that replaces the RGB input to MAMNet. After every optimizer.step() each kernel is projected to zero-mean () to enforce illumination cancellation in the log-domain. A consistency loss with random regularises kernel invariance; an adaptive weight holds the II contribution at of the task loss (gradients flow only through ). IIM adds 0.008 M parameters.
| Hyperparameter | Value |
| IIM module | |
| Num. kernels | 8 |
| Kernel size | |
| Cross-channel pairs | (R,B), (R,G), (G,B) 24 features |
| Zero-mean projection | After every optimizer step |
| Loss | |
| II loss | SmoothL1 () on IIM features |
| II weighting | Adaptive, target ratio |
| Gamma range | |
| Total loss |
Table 12. MAMNet + IIM hyperparameters; settings matching Table 9 are omitted.
MAMNet + ISW
MAMNet + ISW. ISW [23] is a training-only feature-covariance regulariser; the inference architecture is unchanged. Phase 1 (offline): each training image and its photometric augmentation (ColorJitter(0.4,0.4,0.4,0.1) + GaussianBlur p=0.5) are passed through the encoder; the per-instance covariance of standardised features at feat1, feat2, feat3 is computed, and the variance of each covariance element across the original/augmented pair is averaged over the training set. Upper-triangular variance values are partitioned by -means (); the top clusters are flagged as style-sensitive, producing a binary mask per layer. Phase 2 (training): at each step, forward hooks capture features at the same three layers; per-image standardised covariance is masked element-wise with and reduced to (per-image mean), then averaged across the three layers. Total loss is .
| Hyperparameter | Value |
| ISW configuration | |
| (ISW weight) | 0.6 |
| Hooked layers | feat1, feat2, feat3 |
| -means / | 3 / 1 (bottom 1 cluster = content) |
| Photometric aug. (mask precomp.) | ColorJitter(0.4,0.4,0.4,0.1) + GaussianBlur =0.5 |
| Loss reduction | Per-image mean, then layer mean |
| Total loss |
Table 13. MAMNet + ISW hyperparameters; settings matching Table 9 are omitted.
MAMNet + MRFP+
MAMNet + MRFP+. MRFP+ [25] adds three training-only stochastic perturbations. HRFP attaches after conv1-bn1-relu and before layer1 (64 ch, ): a 4-encoder/4-decoder overcomplete autoencoder with 1.2 bilinear upsampling per encoder layer; its conv weights are He-reinitialised and BN , resampled from at every forward pass, and its output is added to the encoder stream under no_grad+detach. NP+ attaches at feat4 (5124848): channel-mean perturbation with and a normalised cross-batch variance. HRFP+ adds the max-resolution intermediate feature from the HRFP encoder to decoder stage dec3; three parameter-free InstanceNorm2d layers (affine=False) are applied on feat1/feat2/feat3. Each module is toggled independently with =0.5. At inference all modules are disabled and the model is identical to base MAMNet (HRFP weights add 74 K frozen parameters never saved).
| Hyperparameter | Value |
| Perturbation modules (training only) | |
| HRFP attach point | After stem (64 ch, ) |
| NP+ attach point | feat4 (512 ch, ) |
| HRFP+ attach point | Decoder dec3 (64 ch) |
| Inst. Norm layers | feat1, feat2, feat3 (affine=False) |
| HRFP / NP+ / HRFP+ prob. | 0.5 / 0.5 / 0.5 (independent) |
| HRFP arch. | 4 enc + 4 dec, 1.2 scale per layer |
| HRFP BN std | 0.5 |
| HRFP weights | Frozen, re-randomised every forward pass |
Table 14. MAMNet + MRFP+ hyperparameters; settings matching Table 9 are omitted.
MAMNet + FADA
MAMNet + FADA. FADA [26] freezes the ResNet-34 encoder (requires_grad=False; BN in eval mode using pretrained running statistics) and inserts a frequency adapter at feat3, feat4, and feat5. Each FADA block applies a 2-D Haar DWT ; the LL subband is processed by a low-frequency branch (LoRA tokens with , rank ; scaled dot attention; MLP projection; skip fusion) for content stabilisation; the concatenated LH/HL/HH are processed by a high-frequency branch (separate LoRA tokens ; scaled-dot attention; Instance Normalisation applied to the similarity map; MLP projection; skip fusion) for style suppression; the four adapted subbands are recombined by inverse Haar DWT. Stages 1 and 2 are excluded for memory reasons (spatial dims and yield prohibitive token attention). Trainable parameters: 8.4 M (FADA 2.5 M, MSCAF + decoder + AUX 5.9 M); frozen 21.3 M.
| Hyperparameter | Value |
| FADA configuration | |
| Adapter stages | feat3, feat4, feat5 |
| LoRA rank | 16 |
| Token length | 100 |
| LF branch | LoRA tokens on LL (content) |
| HF branch | LoRA tokens + IN on similarity map (LH/HL/HH) |
| Decomposition | Haar DWT (orthogonal, invertible) |
| Optimisation overrides | |
| Backbone state | Frozen (BN in eval mode) |
| Trainable groups | FADA blocks + MSCAF + decoder + AUX |
Table 15. MAMNet + FADA hyperparameters; settings matching Table 9 are omitted.
OGLANet
Architecture. OGLANet [17] is a four-stage encoder–decoder with dense skip connections. The encoder consists of five Global-Local Aware Module (GLAM) stages on top of a ResNet-34 backbone (pretrained on ImageNet via torchvision.models.resnet34); the original paper specifies ResNet-101, but ResNet-34 is substituted to match MAMNet’s parameter budget. Each GLAM wraps one ResNet layer (LFEM) between two GFEM self-attention modules ( Q/K/V projections, sigmoid-gated, followed by conv with stride 2 for downsampling or stride 1 for refinement). GLAM5 is an additional stage (not in the original figure) that expands channels with a stride-2 conv plus a non-downsampling GFEM. The DFFM forms four decoder maps – by concatenating multi-scale encoder features iteratively downsampled with stride-2 convs, each followed by a channel-reduction conv (BN-ReLU). The decoder bilinearly upsamples DFFM features through Conv3×3-BN-ReLU blocks, producing outputs at . The OAM produces six -conv predictions: – from the four decoder scales, from concat(), from concat(). All six are deeply supervised at training; only is used at inference. Input is (RGBC); the contrast-channel conv1 weight is initialised as the channel-mean of the pretrained RGB weights. Output stride 1.
DFFM reimplementation defect. We reimplemented the DFFM from the published architecture figure, and the reimplementation makes OGLANet inference depend on the random state and the batch size. Its 20 downsampling convolutions and their batch-normalization layers are constructed inside forward() on every call and never registered as modules. They are absent from the checkpoint, receive fresh random weights on every forward pass, and stay in training mode under model.eval(), so they normalize with the statistics of the current batch. With a fixed seed, inference is bit-identical across jobs and nodes. Over repeated unseeded passes of one checkpoint on the same 150 test images, the standard deviation of the cell mean reaches 0.42 mIoU. Changing the inference seed among three values moves the cell mean by up to 0.66 mIoU on an OGLANet-SIB checkpoint and by up to 1.37 mIoU on a published Vanilla checkpoint, both on Chicago at high resolution. Every OGLANet value in the paper is one inference pass at its run’s inference batch size, and its confidence interval does not include this variation. We left the module unchanged because every reported OGLANet number was produced with it, and the released code contains the same module.
OGLANet + FDA
OGLANet + FDA. FDA is applied identically to MAMNet + FDA (see § E.2): per-channel 2-D FFT, central amplitude swap with , source phase preserved, reconstruction clipped to . Applied as input-level preprocessing in the dataset pipeline before standard augmentation; model architecture and optimisation are unchanged. Run in LOCO mode only; the target pool is sampled uniformly per iteration from the held-out city’s train/images.
OGLANet + SegDeSiC
OGLANet + SegDeSiC. The SegDeSiC module is attached in parallel to feat5 (, GLAMEncoder output): global average pool Linear(1024, 256) ReLU+Dropout(0.3) Linear(256, 256) ReLU+Dropout(0.3) Linear(256, 40) normalisation; weights Xavier uniform. GRID encoding (, ) and the cosine-dissimilarity domain loss are identical to MAMNet + SegDeSiC (§E.3). Total loss combines the deep-supervision six-CE segmentation loss (equal weights ) with source and target geographic terms scaled by (target term LOCO-only). Coordinates are not supplied at validation or test time.
| Hyperparameter | Value |
| FDA low-freq. ratio | 0.01 |
| Target pool size | images |
| Target domain | Held-out city train/images (unlabelled) |
| Target sampling | Uniform random per iteration |
| Application point | Input (dataset, pre-augmentation) |
| Modes run | LOCO only |
Table 17. OGLANet + FDA hyperparameters; settings matching Table 16 are omitted.
| Hyperparameter | Value |
| SegDeSiC module | |
| Attachment point | feat5 () |
| GRID scales / range | 10 / log-spaced |
| Encoding dim | 40 |
| MLP hidden dim / dropout | 256 / 0.3 |
| Loss | |
| Seg. loss | Mean CE over –, equal weights |
| Domain loss | , batch mean |
| 0.5 (both source and target terms) | |
| Target domain (LOCO) | Held-out city train split (unlabelled) |
| Optimisation overrides | |
| Learning rate | (vs. base) |
| Batch size | 4 (vs. 8 base) |
| Modes run | LOCO only |
Table 18. OGLANet + SegDeSiC hyperparameters; settings matching Table 16 are omitted.
OGLANet + IIM
OGLANet + IIM. The IIM preprocessor (identical to §E.4: 8 learnable zero-mean kernels, three cross-channel pairs, FuseConv, adaptive II loss with and ) is inserted before the OGLANet stem. The downstream architecture (GLAMEncoder, DFFM, decoder, OAM) is unchanged. The optimiser is switched from Adamax to Adam to handle the gradient-scale gap between the small IIM module ( M) and the pretrained ResNet-34 backbone ( M).
| Hyperparameter | Value |
| IIM module | |
| Num. kernels | 8 |
| Kernel size | |
| Cross-channel pairs | (R,B), (R,G), (G,B) 24 features |
| Zero-mean projection | After every optimizer step |
| Loss | |
| II loss | SmoothL1 () on IIM features |
| II weighting | Adaptive, target ratio |
| Gamma range | |
| Optimisation overrides | |
| Optimizer | Adam (vs. Adamax) |
| Learning rate | (vs. ) |
| Weight decay |
Table 19. OGLANet + IIM hyperparameters; settings matching Table 16 are omitted.
OGLANet + ISW
OGLANet + ISW. ISW is applied as in §E.5 with hooks on glam1, glam2, glam3 (the first three GLAMEncoder outputs). Phase 1 precomputes binary masks once per LOCO fold from the photometric augmentation covariance variance, partitioned by -means (, ); Phase 2 adds to the six-CE segmentation loss, with averaged per-image and across the three layers. Inference is unchanged.
| Hyperparameter | Value |
| (ISW weight) | 0.6 |
| Hooked layers | glam1, glam2, glam3 |
| -means / | 3 / 1 |
| Photometric aug. (mask precomp.) | ColorJitter(0.4,0.4,0.4,0.1) + GaussianBlur |
| Loss reduction | Per-image mean, then layer mean |
Table 20. OGLANet + ISW hyperparameters; settings matching Table 16 are omitted.
OGLANet + MRFP+
OGLANet + MRFP+. HRFP attaches after conv1-bn1-relu of the GLAMEncoder stem (64 ch, ); NP+ attaches at glam4 (512 ch, ); HRFP+ adds the max-resolution intermediate HRFP feature to the decoder’s s1_d_up stage (64 ch, 384384). Modules are toggled independently with . The InstanceNorm layers specified in the original MRFP+ paper are removed (Option A): the GLAM stages end with two GFEM self-attention modules whose channel-wise statistics are essential to subsequent processing, and applying InstanceNorm2d (affine=False) after GFEM caused class collapse (mIoU ) in all three cities. HRFP and NP+ alone span both ends of the perturbation spectrum. HRFP weights ( K) are frozen, re-randomised each forward pass, and discarded at inference.
Table 21: OGLANet + MRFP+ hyperparameters; settings matching Table 16 are omitted.
| Hyperparameter | Value |
| Perturbation modules (training only) | |
| HRFP attach point | After stem (64 ch, ) |
| NP+ attach point | glam4 (512 ch, ) |
| HRFP+ attach point | Decoder s1_d_up (64 ch, ) |
| Inst. Norm layers | Removed (Option A) |
| HRFP / NP+ / HRFP+ prob. | 0.5 / 0.5 / 0.5 |
| HRFP arch. | 4 enc + 4 dec, scale per layer |
| HRFP BN std | 0.5 |
| Optimisation overrides | |
| Learning rate | (vs. ) |
| Weight decay | |
| Loss reduction | Per-image mean CE over – |
Table 21.
OGLANet + FADA
OGLANet + FADA. Only the ResNet-34 core inside each GLAM stage (conv1, bn1, layer1–layer4) is frozen; the GFEM modules, glam5_conv, glam5_gfem, DFFM, decoder, and OAM remain trainable because they are not pretrained and freezing them collapsed mIoU to in pilot runs. FADA blocks are inserted between the frozen LFEM output and the trainable GFEM at stages 3 and 4 (256 ch and 512 ch, and ); stages 1, 2, and 5 are excluded for either spatial-cost or non-pretrained-component reasons. The block internals (Haar DWT, LL LoRA-token branch, IN-on-similarity-map HF branch, inverse Haar DWT) are identical to MAMNet + FADA (§E.7). Trainable parameters: – M; frozen M.
Table 22: OGLANet + FADA hyperparameters; settings matching Table 16 are omitted.
| Hyperparameter | Value |
| FADA configuration | |
| Adapter stages | GLAM 3, 4 (LFEM output, pre-GFEM) |
| LoRA rank | 16 |
| Token length | 100 |
| Decomposition | Haar DWT |
| Optimisation overrides | |
| Optimizer | Adam (vs. Adamax) |
| Learning rate | (vs. ) |
| Weight decay | |
| LR scheduler | ReduceLROnPlateau (factor=0.5, patience=3) |
| Frozen | ResNet-34 core only (BN in eval mode) |
| Trainable | FADA + GFEM + DFFM + decoder + OAM |
Table 22.
DINOv3
Architecture. DINOv3 ViT-S/16 backbone (dinov3_vits16): patch size , embedding dimension , 12 transformer blocks. Pre-trained weights dinov3_vits16_pretrain_lvd1689m-08c60483.pth (github.com/facebookresearch/dinov3); all backbone parameters trainable (frozen_stages=-1). Intermediate features extracted after blocks 3, 6, 9, and 11 at , concatenated along the channel axis, and fused by a -Conv-BN-ReLU (, bias=False). The decoder applies four ConvTranspose2d (, stride 2) stages () each followed by Conv-BN-ReLU; channel sequence [384, 256, 128, 64, 32]. No encoder skip connections (all ViT features share the same spatial resolution); a final conv yields 2-class logits. No auxiliary branches; single output head. Input is RGB, yielding exactly patches with no padding; output stride 16.
DINOv3 + FDA
DINOv3 + FDA. FDA is applied as in §E.2 (per-channel 2-D FFT, central amplitude swap with , source phase preserved, reconstruction clipped). Inserted in the dataset pipeline before PILtensor and ImageNet normalisation; validation and test images are unmodified. FDA and SegDeSiC are mutually exclusive in all launch scripts. Run in LOCO mode only.
| Hyperparameter | Value |
| FDA low-freq. ratio | 0.01 |
| Target pool size | images |
| Target domain | Held-out city train/images (unlabelled) |
| Application point | Input (pre-tensor, pre-normalisation) |
| Early-stopping patience | 10 (all modes) |
| Modes run | LOCO only |
Table 24. DINOv3 + FDA hyperparameters; settings matching Table 23 are omitted.
DINOv3 + SegDeSiC
DINOv3 + SegDeSiC. A SegDesicModule is attached to feat_block11 (, the deepest extracted ViT feature). The module predicts a 40-d GRID encoding (, ) via global average pool and a 256-dim hidden MLP, with cosine-dissimilarity loss against the ground-truth encoding from . In LOCO mode an additional unlabelled forward pass over the held-out city’s training images contributes a target-domain cosine loss; no segmentation loss is applied to the target. Total loss: .
DINOv3 + IIM
DINOv3 + IIM. The IIM preprocessor (8 learnable zero-mean kernels, three cross-channel pairs, FuseConv, adaptive II loss with , ; identical to §E.4) is inserted before the ViT-S/16 stem. Output is the 3-channel fused image re-normalised with ImageNet statistics. The base DINOv3 strict mIoU decision metric is replaced by tolerant per-image mIoU ( px) for consistency with the IIM ablation protocol.
Table 25: DINOv3 + SegDeSiC hyperparameters; settings matching Table 23 are omitted.
| Hyperparameter | Value |
| SegDeSiC module | |
| Attachment point | feat_block11 (384 × 24 × 24) |
| GRID scales / range | 10 / [, ] |
| Encoding dim | 40 |
| MLP hidden dim / dropout | 256 / 0.3 |
| Loss | |
| Domain loss | |
| 0.5 (both terms) | |
| Auxiliary weight | 0 (no aux. branches) |
| Target domain (LOCO) | Held-out city train split (unlabelled) |
| Optimisation overrides | |
| Learning rate | (vs. ) |
| Batch size | 4 (vs. 8) |
| Early-stopping patience | 10 (all modes) |
Table 25.
Table 26: DINOv3 + IIM hyperparameters; settings matching Table 23 are omitted.
| Hyperparameter | Value |
| IIM module | |
| Num. kernels | 8 |
| Kernel size | |
| Cross-channel pairs | (R,B), (R,G), (G,B) → 24 features |
| Zero-mean projection | After every optimizer step |
| Loss | |
| II loss | SmoothL1 () |
| II weighting | Adaptive, |
| Gamma range | |
| Optimisation overrides | |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early-stopping patience | 15 epochs |
Table 26.
DINOv3 + ISW
DINOv3 + ISW. ISW hooks are registered on the outputs of ViT blocks 3, 6, and 9 (block 11, used by the decoder, is left untouched). The token output is stripped of the CLS token and the 576 patch tokens reshaped to before the standardised covariance is computed. Mask precomputation and loss formulation are identical to §E.5. Inference is unchanged.
Table 27: DINOv3 + ISW hyperparameters; settings matching Table 23 are omitted.
| Hyperparameter | Value |
| (ISW weight) | 0.6 |
| Hooked blocks | ViT blocks 3, 6, 9 |
| Token reshape | Strip CLS; 576 → |
| -means / | 3 / 1 |
| Photometric aug. (mask precomp.) | ColorJitter(0.4,0.4,0.4,0.1) + GaussianBlur |
| Loss reduction | Per-image mean, then block mean |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early-stopping patience | 15 epochs |
Table 27.
DINOv3 + MRFP+
DINOv3 + MRFP+. HRFP attaches at feat_block3 (); NP+ attaches at feat_block11 (); HRFP+ projects the max-resolution intermediate HRFP feature (, 384 ch) through a frozen conv () and adds it to the decoder’s up3 output (64 ch, ). Modules are toggled independently with . As in §E.13, the original-paper InstanceNorm layers are removed: in DINOv3 the four block features feed a single Conv-BatchNorm2d-ReLU fusion, and applying unconditional IN to blocks 6 and 9 during training corrupted that BN’s running statistics, producing decoder checkerboard artefacts and zero mIoU at evaluation. HRFP and NP+ alone span both ends of the perturbation spectrum. HRFP+HRFP+ projection adds K frozen parameters; inference is identical to base DINOv3.
| Hyperparameter | Value |
| Perturbation modules (training only) | |
| HRFP attach point | feat_block3 (384 ch, ) |
| NP+ attach point | feat_block11 (384 ch, ) |
| HRFP+ attach point | Decoder up3 (64 ch, ) |
| HRFP+ projection | Frozen conv |
| Inst. Norm layers | Removed |
| HRFP / NP+ / HRFP+ prob. | 0.5 / 0.5 / 0.5 |
| HRFP arch. | 4 enc + 4 dec, scale per layer |
| HRFP BN std | 0.5 |
| Optimisation overrides | |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early-stopping patience | 15 epochs |
Table 28. DINOv3 + MRFP+ hyperparameters; settings matching Table 23 are omitted.
DINOv3 + FADA
DINOv3+FADA. The DINOv3 ViT-S/16 backbone is fully frozen (requires_grad=False); ViT uses LayerNorm so no eval-mode override is needed. FADA blocks are inserted after ViT blocks 3, 6, 9, and 11 (all four feature-extraction blocks), each operating on the spatial patch tokens [B, 384, 24, 24]: 2-D Haar DWT, LL LoRA-token branch (content), HF LoRA-token branch with Instance Normalisation on the similarity map (style suppression), inverse Haar DWT, and re-injection into the token stream. Gradients propagate through frozen blocks (parameters not updated) to upstream adapters. Trainable parameters: M ( M FADA + M decoder).
| Hyperparameter | Value |
| FADA configuration | |
| Adapter blocks | ViT blocks 3, 6, 9, 11 |
| LoRA rank | 16 |
| Token length | 100 |
| LF branch | LoRA tokens on LL |
| HF branch | LoRA tokens + IN on similarity map (LH/HL/HH) |
| Decomposition | Haar DWT |
| Optimisation overrides | |
| Optimizer | Adam (vs. AdamW) |
| Learning rate | (vs. ) |
| Weight decay | (vs. 0.05) |
| LR scheduler | ReduceLROnPlateau (factor=0.5, patience=3) |
| Backbone state | Frozen (LayerNorm; no eval-mode override) |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early-stopping patience | 15 epochs |
Table 29. DINOv3 + FADA hyperparameters; settings matching Table 23 are omitted.
CrossEarth
CrossEarth. We train CrossEarth [9] from its official release on our splits, starting from the release’s DINOv2 [43] ViT-L/14 backbone initialisation at patch size 16. The backbone is frozen; the trainable components are the Rein adapters [44] (token length 100, LoRA rank 16) and a Mask2Former decoder [45] with a two-class head. The code release omits four components described in the CrossEarth paper (the Geospatial Semantic Extractor, the cross-attention Injector, the masked-image-modelling branch with its ASPP decoder, and Earth-Style Injection), so our cells train the released Rein and Mask2Former configuration without the Million-AID style pool. No CrossEarth per-benchmark checkpoint is used. Table 30 lists the configuration and our deviations from the release, which adapt it to our patches, our 450-image training sets, the 4-hour partition limit, and source-only validation.
| Hyperparameter | Value |
| Architecture | |
| Backbone | DINOv2 ViT-L/14 at patch 16, frozen (303.7 M parameters) |
| Adapters | Rein, token length 100, LoRA rank 16 |
| Decoder | Mask2Former, 2 classes |
| Trainable parameters | 23.6 M |
| Optimisation | |
| Optimizer | AdamW (, , ) |
| Learning rate | |
| Weight decay | 0.05 |
| LR scheduler | Polynomial (power 0.9, horizon 40,000 iterations) |
| Batch size | 4 (release: 1) |
| Iterations | 10,000 (release: 30,000) |
| Crop | (release: ) |
| Precision | bf16 autocast; deformable attention in fp32 |
| Validation | Every 500 iterations, source cities only |
| Checkpoint selection | Best two-class tolerant mIoU ( px) on source-city validation |
| Inference | Whole image (release: sliding window) |
| Infrastructure | |
| Hardware | 1 NVIDIA A100 40 GB (NCSA Delta, gpuA100x4) |
Table 30. CrossEarth configuration. Every CrossEarth cell (Vanilla, FDA, SIB, BSM; LOCO and within-city upper bound) uses these settings, high resolution only, seed 0.
CrossEarth + FDA. FDA is applied as for the other architectures: low-frequency ratio ( swapped region per channel), a pool of 100 unlabelled images from the held-out city’s train/images, input stage, training only.
CrossEarth + SIB. SIB on CrossEarth is the Haar and VIB path of Appendix § 17; FDA preprocessing and skip-attention gates are CNN-only components and are not applied. One shared-weight SIB block is inserted before the Mask2Former pixel decoder and applied to all four backbone feature scales (strides 4, 8, 16, 32), adding 2.4 M trainable parameters. The VIB weights, hidden widths ( and at ), KL warm-up and coverage criterion follow Appendix § 17; is fitted on source-city validation per cell.
BSM
BSM. No official BSM code has been released, so we reimplemented it from the paper [22]. At each training step a frozen AdaIN style-transfer network [46] re-renders every image in the batch with the style of another image from the same batch (a random permutation with distinct pairs), and the result is mixed with the original at a fixed coefficient . The style source is the training batch itself, so no external or held-out-city imagery enters training. BSM is applied to the RGB input before the network, with the architecture’s input normalisation inverted before mixing and reapplied after; the contrast channel of MAMNet and OGLANet passes through unmodified. BSM is disabled at evaluation. Each architecture otherwise keeps its base configuration (Table 9, Table 16, Table 23, Table 30). The DINOv3/Chicago cell collapsed at seed 0 (best source-city validation mIoU 5.51, below the training floor of 10.0) and is excluded from Table 2 without imputation.
Table 9: MAMNet training configuration. †Shell script overrides the Python-file default of .
| Hyperparameter | Value |
| Optimisation | |
| Optimizer | Adam |
| Learning rate | † |
| Weight decay | |
| Batch size | 8 |
| Epochs | 100 |
| LR scheduler | ReduceLROnPlateau (mode=max, factor=0.5, patience=3) |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early stopping | patience = 10 epochs |
| Loss | |
| Main loss | nn.CrossEntropyLoss (2-class, unweighted) |
| Total loss | |
| Data | |
| Input size | |
| Channels | 4 (RGBC) |
| Train aug. | H-flip , V-flip |
| Val/test aug. | None (resize + ImageNet normalisation only) |
| Post-processing | Remove connected components px |
| Infrastructure | |
| Hardware | NVIDIA A100 (NCSA Delta, gpuA100x4) |
Table 9.
| Hyperparameter | Value |
| Optimisation | |
| Optimizer | Adamax |
| Learning rate | |
| Weight decay | Adamax default |
| Batch size | 8 (script override; train.py default 12) |
| Epochs | 100 |
| LR scheduler | ReduceLROnPlateau (mode=max, factor=0.5, patience=5) |
| Decision metric | Per-image tolerant mIoU ( px) |
| Early stopping | patience = 10 epochs |
| Loss | |
| Loss | OGLANetLoss: six CE terms (– on –), equal weights |
| Post-processing | Remove connected components < 10 px |
| Data | |
| Input | (RGBC) |
| Infrastructure | |
| Hardware | NVIDIA A100 40 GB (NCSA Delta, gpuA100x4) |
| Mixed precision | Disabled |
| Gradient clip. | Disabled |
Table 16. OGLANet training configuration.
| Hyperparameter | Value |
| Optimisation | |
| Optimizer | AdamW (, ) |
| Learning rate | (CLI; argparse default ) |
| Weight decay | 0.05 |
| Batch size | 8 |
| Epochs (max) | 100 |
| LR scheduler | Cosine + linear warmup (5 ep., ) |
| Decision metric | Per-image strict mIoU (DetailedEvaluator) |
| Early stopping | 20 (single, LOCO); 10 (all) |
| Loss | |
| Loss | nn.CrossEntropyLoss (2-class, no weights, no aux.) |
| Post-processing | Remove connected components px |
| Data | |
| Input size | |
| Channels | 3 (RGB) |
| Train aug. | H-flip , V-flip (PIL, train split) |
| Normalisation | ImageNet , |
| Infrastructure | |
| Hardware | A100 (NCSA Delta, gpuA100x4) |
| Sampling seed | 42 (dataset subsampling only) |
Table 23. DINOv3 training configuration.
| Hyperparameter | Value |
| Style network | AdaIN: VGG-19 encoder to relu4_1 + mirror decoder, frozen |
| Style-network weights | Public pretrained AdaIN release; 0 trainable parameters |
| Style source | Permuted training batch |
| Mixing coefficient | 0.5 (fixed; checked for non-degeneracy on source-city validation) |
| Application probability | 1.0 per training batch |
| Batch size | 8 (CrossEarth: 4) |
| Early-stopping patience | 20 epochs (vs. 10; CrossEarth: fixed 10,000 iterations) |
| Modes run | LOCO, high resolution only |
Table 31. BSM hyperparameters; settings matching each architecture’s base configuration are omitted.
SAR
SAR. SAR [28] trains nothing. It adapts the normalisation-layer affine parameters of each cell’s Vanilla LOCO checkpoint (for MAMNet, the matched-450 retrain) on each test image’s own prediction, by minimising the entropy of reliable pixels under sharpness-aware minimisation [47]. We run it episodically at batch size 1: parameters are restored before every image, so nothing accumulates across the target set, and no held-out-city label or source image is used. Normalisation layers stay in evaluation mode on source running statistics, because MAMNet and OGLANet return auxiliary and deep-supervision outputs in training mode. We keep upstream’s skip list and collect the affine parameters of every BatchNorm, LayerNorm and GroupNorm layer. The two-class entropy constants are upstream’s, rescaled from to and never tuned on target mIoU. When the reliability filter admits no pixel, the step is skipped. Each SAR row in Table 2 is the Vanilla cell’s per-image values plus the change SAR produces within its own pass (final step minus step 0), so the row isolates the adaptation effect from any difference between the SAR pass and the stored Vanilla inference.
| Hyperparameter | Value |
| Adapted parameters | Affine of BatchNorm/LayerNorm/GroupNorm layers |
| Excluded (upstream) | layer4, ViT blocks 9–11, final norm |
| Normalisation statistics | Source running statistics (evaluation mode) |
| Optimizer | SAM () over SGD (momentum 0.9) |
| Learning rate | (upstream ViT default, not batch-scaled) |
| Steps per image | 1 (MAMNet, OGLANet, DINOv3); 10 (CrossEarth) |
| Entropy margin | |
| Model-recovery threshold | 0.0201 nats |
| EMA momentum | 0.9 |
| Batch size | 1, episodic |
| Hardware | NVIDIA A100 40 GB (NCSA Delta, gpuA100x4) |
Table 32. SAR hyperparameters.
Selection. For MAMNet, OGLANet and DINOv3, we selected the learning rate ( the base rate) and step count () on source-city validation in a pilot on the Phoenix fold that pooled DINOv3 and MAMNet. Upstream additionally scales the learning rate by batch size over 64, which at batch size 1 gives ; we do not apply that scaling.
SAR on CrossEarth. CrossEarth exposes no logits, so SAR minimises the entropy of the logarithm of the normalised two-class Mask2Former score. Its recipe (base rate, 10 steps) is borrowed from the S-EO pilot, which selected on S-EO source-city validation with MAMNet, OGLANet and DINOv3 and did not include CrossEarth; no internal CrossEarth cell selected it. The SAR pass runs in fp32 while the parent cells were inferred under bf16 autocast. Before adaptation, the SAR pass differs from the stored Vanilla value by +0.0063 (Chicago), +0.1053 (Miami) and +0.0070 (Phoenix) mIoU; most of the Miami difference comes from one image with 46 foreground pixels. This difference does not enter the row, which adds only the within-pass change: −0.0011, −0.0054 and +0.0020 mIoU. The run’s divergence guard, which fires when more than 20% of images return a foreground fraction below 0.01, fires on Miami and Phoenix. The number of such images is the same before and after adaptation in every city (11, 42 and 53 of 150), so the flag reflects the parent model.
Full Per-Cell LOCO Results
This appendix reports the full per-cell breakdown of the 158 LOCO test cells summarized in Table 2. Table 33–Table 35 cover the 126 cells of MAMNet, OGLANet and DINOv3 under Vanilla and the six training-time methods, and Table 36 the 32 high-resolution SAR, BSM and CrossEarth cells. Each table reports, for each (architecture, method, held-out city, native resolution) cell, the within-city upper-bound mIoU, the LOCO mIoU, the per-image mean to the upper bound, the 95% paired-bootstrap confidence interval on , and the two-tailed bootstrap -value ( resamples on per-image values). All values are two-class boundary-tolerant mIoU ( px, Appendix § 12). indicates the bootstrap resolution floor; cells marked † are not significant at . Table 37–Table 39 report pooled summaries of the 126 cells by held-out city, native resolution, and method.
| Method | City | Res. | Upper | LOCO | 95% CI | ||
| Vanilla | Chicago | HR | 86.31 | 80.45 | |||
| Chicago | MR | 86.35 | 83.22 | ||||
| Miami | HR | 79.11 | 74.67 | ||||
| Miami | MR | 78.89 | 76.75 | ||||
| Phoenix | HR | 73.75 | 64.23 | ||||
| Phoenix | MR | 85.64 | 83.40 | ||||
| FDA | Chicago | HR | 86.31 | 79.77 | |||
| Chicago | MR | 86.35 | 82.88 | ||||
| Miami | HR | 79.11 | 73.10 | ||||
| Miami | MR | 78.89 | 75.71 | ||||
| Phoenix | HR | 73.75 | 67.89 | ||||
| Phoenix | MR | 85.64 | 82.64 | ||||
| SegDeSiC | Chicago | HR | 86.31 | 82.01 | |||
| Chicago | MR | 86.35 | 81.50 | ||||
| Miami | HR | 79.11 | 70.15 | ||||
| Miami | MR | 78.89 | 75.99 | ||||
| Phoenix | HR | 73.75 | 63.54 | ||||
| Phoenix | MR | 85.64 | 83.49 | ||||
| IIM | Chicago | HR | 86.31 | 79.25 | |||
| Chicago | MR | 86.35 | 82.31 | ||||
| Miami | HR | 79.11 | 67.95 | ||||
| Miami | MR | 78.89 | 74.52 | ||||
| Phoenix | HR | 73.75 | 49.69 | ||||
| Phoenix | MR | 85.64 | 77.85 | ||||
| MRFP+ | Chicago | HR | 86.31 | 74.81 | |||
| Chicago | MR | 86.35 | 81.29 | ||||
| Miami | HR | 79.11 | 70.87 | ||||
| Miami | MR | 78.89 | 67.09 | ||||
| Phoenix | HR | 73.75 | 55.16 | ||||
| Phoenix | MR | 85.64 | 78.94 | ||||
| FADA | Chicago | HR | 86.31 | 77.89 | |||
| Chicago | MR | 86.35 | 80.03 | ||||
| Miami | HR | 79.11 | 65.27 | ||||
| Miami | MR | 78.89 | 72.35 | ||||
| Phoenix | HR | 73.75 | 61.79 | ||||
| Phoenix | MR | 85.64 | 76.41 | ||||
| ISW | Chicago | HR | 86.31 | 74.65 | |||
| Chicago | MR | 86.35 | 71.26 | ||||
| Miami | HR | 79.11 | 68.04 | ||||
| Miami | MR | 78.89 | 69.34 | ||||
| Phoenix | HR | 73.75 | 50.47 | ||||
| Phoenix | MR | 85.64 | 76.04 |
Table 33. Full per-cell LOCO results: MAMNet. paired test images per cell. Upper, LOCO in mIoU (%); in mIoU points; CI is the 95% paired-bootstrap interval on . : not significant at .
| Method | City | Res. | Upper | LOCO | 95% CI | ||
| Vanilla | Chicago | HR | 80.27 | 67.26 | −13.01 | [−14.37, −11.67] | < .0002 |
| Chicago | MR | 83.63 | 75.23 | −8.40 | [−9.71, −7.17] | < .0002 | |
| Miami | HR | 76.00 | 72.61 | −3.39 | [−5.36, −1.67] | < .0002 | |
| Miami | MR | 75.76 | 68.11 | −7.65 | [−9.24, −6.16] | < .0002 | |
| Phoenix | HR | 67.82 | 65.24 | −2.58 | [−4.30, −0.88] | .0028 | |
| Phoenix | MR | 81.85 | 71.52 | −10.33 | [−12.00, −8.61] | < .0002 | |
| FDA | Chicago | HR | 80.27 | 72.26 | −8.01 | [−9.15, −6.92] | < .0002 |
| Chicago | MR | 83.63 | 75.70 | −7.94 | [−8.85, −7.04] | < .0002 | |
| Miami | HR | 76.00 | 70.81 | −5.19 | [−6.30, −4.04] | < .0002 | |
| Miami | MR | 75.76 | 71.95 | −3.81 | [−5.15, −2.62] | < .0002 | |
| Phoenix | HR | 67.82 | 58.72 | −9.10 | [−11.21, −6.97] | < .0002 | |
| Phoenix | MR | 81.85 | 67.01 | −14.84 | [−16.56, −13.13] | < .0002 | |
| SegDeSiC | Chicago | HR | 80.27 | 75.52 | −4.75 | [−6.10, −3.35] | < .0002 |
| Chicago | MR | 83.63 | 77.30 | −6.33 | [−7.44, −5.32] | < .0002 | |
| Miami | HR | 76.00 | 71.18 | −4.82 | [−6.76, −3.10] | < .0002 | |
| Miami | MR | 75.76 | 70.72 | −5.04 | [−6.60, −3.63] | < .0002 | |
| Phoenix | HR | 67.82 | 68.04 | +0.22 | [−1.22, +1.66] | .7644† | |
| Phoenix | MR | 81.85 | 75.13 | −6.72 | [−8.10, −5.37] | < .0002 | |
| IIM | Chicago | HR | 80.27 | 77.43 | −2.85 | [−4.04, −1.69] | < .0002 |
| Chicago | MR | 83.63 | 73.58 | −10.05 | [−11.68, −8.54] | < .0002 | |
| Miami | HR | 76.00 | 72.23 | −3.78 | [−5.75, −2.06] | < .0002 | |
| Miami | MR | 75.76 | 69.42 | −6.34 | [−7.98, −4.79] | < .0002 | |
| Phoenix | HR | 67.82 | 56.51 | −11.31 | [−13.57, −9.01] | < .0002 | |
| Phoenix | MR | 81.85 | 79.52 | −2.33 | [−3.79, −0.86] | .0034 | |
| MRFP+ | Chicago | HR | 80.27 | 77.46 | −2.81 | [−3.67, −1.94] | < .0002 |
| Chicago | MR | 83.63 | 78.92 | −4.71 | [−5.72, −3.78] | < .0002 | |
| Miami | HR | 76.00 | 73.20 | −2.80 | [−4.13, −1.52] | < .0002 | |
| Miami | MR | 75.76 | 72.21 | −3.55 | [−4.98, −2.17] | < .0002 | |
| Phoenix | HR | 67.82 | 60.78 | −7.04 | [−9.30, −4.72] | < .0002 | |
| Phoenix | MR | 81.85 | 73.47 | −8.37 | [−9.99, −6.84] | < .0002 | |
| FADA | Chicago | HR | 80.27 | 70.15 | −10.12 | [−11.56, −8.62] | < .0002 |
| Chicago | MR | 83.63 | 63.10 | −20.54 | [−22.32, −18.83] | < .0002 | |
| Miami | HR | 76.00 | 71.43 | −4.57 | [−5.64, −3.59] | < .0002 | |
| Miami | MR | 75.76 | 66.73 | −9.03 | [−10.80, −7.36] | < .0002 | |
| Phoenix | HR | 67.82 | 62.81 | −5.01 | [−6.66, −3.39] | < .0002 | |
| Phoenix | MR | 81.85 | 67.56 | −14.29 | [−16.02, −12.60] | < .0002 | |
| ISW | Chicago | HR | 80.27 | 57.60 | −22.68 | [−24.35, −20.92] | < .0002 |
| Chicago | MR | 83.63 | 61.66 | −21.98 | [−23.39, −20.55] | < .0002 | |
| Miami | HR | 76.00 | 59.86 | −16.15 | [−18.16, −14.19] | < .0002 | |
| Miami | MR | 75.76 | 58.01 | −17.75 | [−19.91, −15.59] | < .0002 | |
| Phoenix | HR | 67.82 | 55.68 | −12.14 | [−13.87, −10.43] | < .0002 | |
| Phoenix | MR | 81.85 | 61.41 | −20.44 | [−21.91, −18.90] | < .0002 |
Table 35. Full per-cell LOCO results: DINOv3. Conventions as in Table 33.
| Arch. | Method | City | Upper | LOCO | 95% CI | ||
| MAMNet | SAR | Chicago | 86.31 | 80.45 | −5.87 | [−6.89, −4.88] | <.0002 |
| Miami | 79.11 | 74.66 | −4.45 | [−6.05, −2.93] | <.0002 | ||
| Phoenix | 73.75 | 64.22 | −9.53 | [−11.51, −7.54] | <.0002 | ||
| OGLANet | SAR | Chicago | 79.12 | 68.70 | −10.42 | [−11.99, −8.79] | <.0002 |
| Miami | 73.89 | 66.36 | −7.53 | [−10.00, −5.13] | <.0002 | ||
| Phoenix | 66.15 | 61.18 | −4.98 | [−6.28, −3.66] | <.0002 | ||
| DINOv3 | SAR | Chicago | 80.27 | 67.25 | −13.02 | [−14.38, −11.67] | <.0002 |
| Miami | 76.00 | 72.61 | −3.39 | [−5.36, −1.67] | <.0002 | ||
| Phoenix | 67.82 | 65.24 | −2.58 | [−4.30, −0.88] | .0028 | ||
| MAMNet | BSM | Chicago | 86.31 | 78.09 | −8.22 | [−9.37, −7.08] | <.0002 |
| Miami | 79.11 | 72.59 | −6.52 | [−8.21, −4.92] | <.0002 | ||
| Phoenix | 73.75 | 67.70 | −6.05 | [−7.74, −4.45] | <.0002 | ||
| OGLANet | BSM | Chicago | 79.12 | 61.73 | −17.40 | [−19.54, −15.30] | <.0002 |
| Miami | 73.89 | 60.65 | −13.24 | [−15.85, −10.70] | <.0002 | ||
| Phoenix | 66.15 | 64.58 | −1.57 | [−2.92, −0.29] | .0156 | ||
| DINOv3 | BSM | Miami | 76.00 | 69.74 | −6.27 | [−8.52, −4.29] | <.0002 |
| Phoenix | 67.82 | 62.66 | −5.16 | [−6.94, −3.36] | <.0002 | ||
| CrossEarth | Vanilla | Chicago | 85.01 | 82.66 | −2.35 | [−3.11, −1.57] | <.0002 |
| Miami | 78.48 | 72.61 | −5.87 | [−7.20, −4.59] | <.0002 | ||
| Phoenix | 72.89 | 70.55 | −2.34 | [−3.53, −1.09] | <.0002 | ||
| CrossEarth | FDA | Chicago | 85.01 | 82.55 | −2.46 | [−3.44, −1.41] | .0002 |
| Miami | 78.48 | 71.49 | −6.99 | [−8.51, −5.55] | <.0002 | ||
| Phoenix | 72.89 | 71.35 | −1.53 | [−2.60, −0.44] | .0068 | ||
| CrossEarth | SIB | Chicago | 85.01 | 79.31 | −5.70 | [−6.54, −4.88] | <.0002 |
| Miami | 78.48 | 73.93 | −4.55 | [−5.72, −3.48] | <.0002 | ||
| Phoenix | 72.89 | 69.80 | −3.08 | [−4.26, −1.90] | <.0002 | ||
| CrossEarth | SAR | Chicago | 85.01 | 82.66 | −2.35 | [−3.11, −1.57] | <.0002 |
| Miami | 78.48 | 72.61 | −5.88 | [−7.21, −4.60] | <.0002 | ||
| Phoenix | 72.89 | 70.55 | −2.34 | [−3.53, −1.09] | <.0002 | ||
| CrossEarth | BSM | Chicago | 85.01 | 81.42 | −3.59 | [−4.41, −2.81] | <.0002 |
| Miami | 78.48 | 74.76 | −3.73 | [−4.67, −2.79] | <.0002 | ||
| Phoenix | 72.89 | 71.06 | −1.83 | [−3.21, −0.40] | .0108 |
Table 36. Full per-cell LOCO results: SAR, BSM and CrossEarth (high resolution). Conventions as in Table 33. CrossEarth cells are measured against CrossEarth’s own upper bound.
| Held-out city | Sig./42 | |||
| Chicago | −11.33 | −2.81 | −28.50 | 42/42 |
| Miami | −7.23 | −2.14 | −17.75 | 42/42 |
| Phoenix | −8.50 | +0.22 | −24.06 | 41/42 |
Table 37. LOCO degradation by held-out city, pooled across architectures and methods ( cells per city).
| Method | ||||
| Vanilla | 18 | |||
| FDA | 18 | |||
| SegDeSiC | 17 | |||
| IIM | 18 | |||
| MRFP+ | 18 | |||
| FADA | 18 | |||
| ISW | 18 |
Table 39. Method-level safety summary, pooled across architectures, cities, and resolutions ( cells per method). Vanilla (no adaptation) is the no-intervention reference. The worst cell is MRFP+ on OGLANet/Chicago at mid resolution () and the best is SegDeSiC on DINOv3/Phoenix at high resolution (, not significant).
Cell provenance and caveats. The high-resolution MAMNet cells and the high-resolution OGLANet SegDeSiC cells are 450-image retrains that match the training budget of the other high-resolution cells; every mid-resolution cell is the published run at its published budget. Each cell is one training seed. OGLANet cells are single inference passes at the published batch size; OGLANet inference is nondeterministic (Appendix § 13.8), with cell means varying by up to 0.42 mIoU over repeated passes, and the intervals do not include this. SAR rows are the Vanilla cell’s per-image values plus SAR’s within-pass change (Appendix § 13.24). BSM DINOv3/Chicago collapsed in training and has no row. CrossEarth cells are compared against CrossEarth’s own upper bound. The intervals in Figure 1 come from the same paired image bootstrap over the 450 high-resolution test images per architecture; the four architectures share bootstrap draws, so overlapping intervals there are not a test between architectures, and no multiplicity correction is applied.
Per-Cell Tables
| Method | City | Res. | Upper | LOCO | 95% CI | ||
| Vanilla | Chicago | HR | 79.12 | 68.47 | −10.66 | [−12.08, −9.21] | <.0002 |
| Chicago | MR | 80.77 | 65.08 | −15.70 | [−17.66, −13.81] | <.0002 | |
| Miami | HR | 73.89 | 66.18 | −7.71 | [−10.20, −5.35] | <.0002 | |
| Miami | MR | 74.19 | 70.73 | −3.45 | [−4.76, −2.24] | <.0002 | |
| Phoenix | HR | 66.15 | 61.18 | −4.98 | [−6.23, −3.72] | <.0002 | |
| Phoenix | MR | 75.39 | 70.32 | −5.07 | [−6.42, −3.76] | <.0002 | |
| FDA | Chicago | HR | 79.12 | 71.07 | −8.05 | [−9.10, −6.97] | <.0002 |
| Chicago | MR | 80.77 | 60.19 | −20.58 | [−22.54, −18.63] | <.0002 | |
| Miami | HR | 73.89 | 67.30 | −6.59 | [−8.34, −4.89] | <.0002 | |
| Miami | MR | 74.19 | 69.82 | −4.36 | [−5.63, −3.17] | <.0002 | |
| Phoenix | HR | 66.15 | 63.27 | −2.88 | [−4.07, −1.69] | <.0002 | |
| Phoenix | MR | 75.39 | 70.09 | −5.30 | [−6.65, −4.01] | <.0002 | |
| SegDeSiC | Chicago | HR | 79.12 | 68.99 | −10.13 | [−11.72, −8.52] | <.0002 |
| Chicago | MR | 80.77 | 59.40 | −21.37 | [−23.64, −19.13] | <.0002 | |
| Miami | HR | 73.89 | 60.69 | −13.20 | [−16.02, −10.49] | <.0002 | |
| Miami | MR | 74.19 | 69.91 | −4.28 | [−5.62, −3.03] | <.0002 | |
| Phoenix | HR | 66.15 | 60.77 | −5.38 | [−6.72, −4.09] | <.0002 | |
| Phoenix | MR | 75.39 | 70.27 | −5.12 | [−6.59, −3.68] | <.0002 | |
| IIM | Chicago | HR | 79.12 | 70.01 | −9.11 | [−10.43, −7.79] | <.0002 |
| Chicago | MR | 80.77 | 62.21 | −18.56 | [−20.84, −16.34] | <.0002 | |
| Miami | HR | 73.89 | 66.36 | −7.53 | [−9.50, −5.61] | <.0002 | |
| Miami | MR | 74.19 | 69.13 | −5.06 | [−6.50, −3.76] | <.0002 | |
| Phoenix | HR | 66.15 | 55.25 | −10.90 | [−12.30, −9.50] | <.0002 | |
| Phoenix | MR | 75.39 | 70.28 | −5.11 | [−6.43, −3.86] | <.0002 | |
| MRFP+ | Chicago | HR | 79.12 | 63.82 | −15.30 | [−16.93, −13.68] | <.0002 |
| Chicago | MR | 80.77 | 52.27 | −28.50 | [−30.86, −26.14] | <.0002 | |
| Miami | HR | 73.89 | 57.30 | −16.59 | [−19.17, −13.99] | <.0002 | |
| Miami | MR | 74.19 | 66.64 | −7.54 | [−9.09, −6.08] | <.0002 | |
| Phoenix | HR | 66.15 | 58.03 | −8.13 | [−9.63, −6.59] | <.0002 | |
| Phoenix | MR | 75.39 | 67.26 | −8.14 | [−9.65, −6.69] | <.0002 | |
| FADA | Chicago | HR | 79.12 | 66.17 | −12.96 | [−14.49, −11.43] | <.0002 |
| Chicago | MR | 80.77 | 53.43 | −27.35 | [−29.43, −25.30] | <.0002 | |
| Miami | HR | 73.89 | 65.20 | −8.69 | [−11.19, −6.19] | <.0002 | |
| Miami | MR | 74.19 | 69.17 | −5.02 | [−6.43, −3.72] | <.0002 | |
| Phoenix | HR | 66.15 | 60.26 | −5.89 | [−7.28, −4.47] | <.0002 | |
| Phoenix | MR | 75.39 | 68.71 | −6.69 | [−8.06, −5.40] | <.0002 | |
| ISW | Chicago | HR | 79.12 | 64.19 | −14.93 | [−16.27, −13.56] | <.0002 |
| Chicago | MR | 80.77 | 59.74 | −21.03 | [−23.13, −18.95] | <.0002 | |
| Miami | HR | 73.89 | 67.78 | −6.11 | [−7.66, −4.57] | <.0002 | |
| Miami | MR | 74.19 | 64.81 | −9.38 | [−10.96, −7.86] | <.0002 | |
| Phoenix | HR | 66.15 | 59.69 | −6.46 | [−8.03, −4.91] | <.0002 | |
| Phoenix | MR | 75.39 | 66.98 | −8.42 | [−9.56, −7.29] | <.0002 |
Table 34. Full per-cell LOCO results: OGLANet. Conventions as in Table 33.
Pooled Summaries
Histogram Matching Intervention
This appendix expands on the pixel-space test reported in § 4.2, which rules out input-preprocessing as a recovery strategy under LOCO.
Procedure. For each (architecture, holdout city) cell under vanilla LOCO, we compute the aggregate intensity histogram (the mean of the three channels) of the source-city training images. We map each test image’s intensity CDF onto this histogram, restore each pixel’s ratio between its channels and its intensity, and run the trained model on the transformed image with no parameter changes. The intervention is parameter-free, applied at test time, and isolates the pixel-statistics hypothesis: if
| Resolution | ||||
| High-res (0.3 m/px) | 62/63 | |||
| Mid-res (0.6 m/px) | 63/63 |
Table 38. LOCO degradation by native resolution, pooled across architectures, methods, and held-out cities ( cells per resolution).
LOCO failure lives at the pixel level, aligning the marginal pixel distribution should recover at least part of the within-city upper bound.
Recovery and inflation. § 4.2 reports the per-architecture median recovery ratio ; all three medians lie in with 95% CIs below zero. At the median, matched images lie 2.5 to 3.7 times as far from the upper bound as the LOCO predictions. False-positive pixel counts rise on 5 of the 9 cells and fall on the other 4. The largest rise is on MAMNet/Chicago, from 201K LOCO false positives to 15.4M after matching (76×). Counts are pooled over the 150 test images with the boundary-tolerant band of the original scorer. excludes images whose upper-minus-LOCO IoU is below 0.02 or undefined, and OGLANet values carry the nondeterminism of Appendix § 13.8.
Implication. The destruction is uniform and large: 0 of 9 cells improve, and the inflation pattern indicates the model’s features are tied to the joint distribution of intensity, texture, and spatial pattern in source-city imagery. Marginal pixel-distribution alignment breaks this coupling without replacing it. Any recovery strategy that operates only on input statistics (gamma correction, contrast normalization, single-channel histogram standardization) faces the same structural problem.
Selective-Prediction Gap: Per-Cell Details
This appendix expands the SP-gap analysis of § 4.3, which characterizes the LOCO failure as class-asymmetric foreground-confidence collapse.
Bootstrap procedure. For each of the 63 LOCO cells (architecture method holdout city, high-resolution only), we compute per-image risk-coverage curves on three pixel populations: foreground (gt_pos, confidence ), background (gt_bg, confidence ), and all pixels (confidence ). The SP gap is ; image-level paired bootstrap () gives 95% CIs on the gap. Cells are paired by image because both upper and LOCO models evaluate on the same 150 test images per cell.
Class-asymmetry distribution. The pooled means reported in § 4.3 (foreground gap , background gap ) reflect a stable per-cell pattern: the foreground gap exceeds the background gap in 55 of 63 cells (sign-test ). The per-method breakdown shown in Figure 3(b) confirms the asymmetry holds for every adaptation method individually; no method produces a foreground gap smaller than its background gap on average.
Logit and ECE corroboration. Two independent per-pixel measurements track the same asymmetry. At the logit level, true-positive logits drop by 0.85 nats on average while false-positive logits drop by 0.15 nats; the asymmetry holds in 55 of 63 cells. Expected calibration error rises by 0.064 on predicted-positive pixels and is unchanged () on predicted-negative pixels; the rise is larger on the predicted-positive side in 44 of 54 cells (nine cells excluded for insufficient predicted-positive coverage on the calibration split). The three measurements (AURC gap, logit drift, ECE rise) are independent in the sense that they depend on different functions of the prediction distribution, and all three identify the foreground-class population as the locus of degradation.
Cross-Entropy AURC Under-Reporting
This appendix details the methodological finding that standard evaluation under-reports the LOCO calibration failure by , motivating the cross-entropy framing in § 4.3.
Why -AURC misses the calibration component. The standard AURC measures per-pixel error: a pixel is correct if its predicted class matches the ground truth, regardless of the predicted probability. Two models that produce identical rank orderings of pixels by confidence yield identical risk-coverage curves, even if one assigns probability to a correct pixel and the other assigns . The -AURC is invariant to any monotonic rescaling of the probability outputs.
Cross-entropy AURC replaces error with per-pixel cross-entropy: on foreground pixels and on background pixels. CE-AURC depends on the probability values themselves, not just rank order, and therefore detects calibration collapse that leaves the rank ordering intact.
Result. The pooled CE-AURC gap exceeds the pooled -AURC gap by a factor of on the foreground population ( against ). The cell-level direction is consistent: 60 of 63 cells have (one-sided paired Wilcoxon ).
Implication. Practitioners who evaluate selective deployment on metrics alone observe less than half of the calibration degradation that cross-entropy scoring measures; the rest is hidden by the rank-invariance of scoring. The under-reporting is not an artifact of the bootstrap or the bin width: it is a structural property of the metric.
Coverage-Threshold Recovery and Calibration
This appendix details the operating-point recovery analysis and the failure of standard post-hoc calibration that together motivate the deployment-time component of SIB (§ 5.1).
Recoverable-threshold definition. For each cell we define , the largest coverage at which the LOCO selective error on the foreground population matches the within-city full-coverage error within a tolerance. Image-level paired bootstrap on the per-image error stack gives a 95% CI on . A finite exists in 59 of 63 cells. The four exceptions are all on Phoenix under aggressive adaptation methods (IIM, ISW, MRFP+); on these cells, no selective-prediction threshold recovers within-city quality.
Required threshold magnitude. Among the 59 recoverable cells, the median corresponds to discarding 25% of foreground pixels. On Phoenix the median is 40%.
Method-ranking stability. We rank the seven methods (Vanilla and six adaptation methods) on the per-method mean foreground SP gap, separately for each of the three architectures, and compute Kendall’s coefficient of concordance across the three rankings: (, chi-square approximation over 7 methods). FDA has the smallest pooled gap () and ISW the largest (), and the pooled ordering is FDA < Vanilla < FDA < SegDeSiC < IIM < MRFP+ < ISW. FDA ranks first on all three architectures, and ISW and MRFP+ each rank between fifth and seventh. On MANet, IIM, ISW and MRFP+ reach , and .
Post-hoc temperature scaling fails under cross-source-city transfer. Standard practice for calibration repair is to fit a single temperature on a labeled validation set. We test this directly: per cell, fit on the source-city validation split (the only labeled data available pre-deployment), apply to the held-out city. The recovery is indistinguishable from no calibration (paired Wilcoxon against zero recovery, ); on 26 of 63 cells, the fitted-on-source temperature actively degrades held-out performance. Coverage thresholding is the only label-free recipe we identify that recovers within-city quality on the held-out city.
INRIA Cross-Task Replication
This appendix details the cross-task replication on INRIA building footprint extraction [33], used in § 4.3 to test whether the SP-gap findings reflect properties of geographic LOCO or of shadow appearance.
Dataset and protocol. INRIA provides 5,000×5,000-pixel aerial tiles at 0.3 m GSD over five cities; we use Austin, Chicago, and Vienna to match the three-city LOCO structure of SHADOWTRANSFER. From each city’s labeled tiles we sample 750 non-overlapping 384 × 384 patches under a fixed seed, split into 450 training, 150 validation, and 150 test images: identical sample sizes to SHADOWTRANSFER’s per-city splits at high resolution. Patches with more than 30% nodata are rejected before sampling. The three LOCO folds (holdout Austin, Chicago, Vienna) follow the same train-on-two/test-on-one structure as § 3.3.
Training matched to shadow. We train MAMNet from scratch using the identical hyperparameters as the shadow MAMNet (Adam lr = , batch 8, 100 epochs, ReduceLROnPlateau with patience 3, early stopping patience 10, ImageNet-pretrained 4-channel ResNet-34 encoder with the contrast channel initialized from RGB mean, MAMNet auxiliary loss with weight 0.4, ±2 px boundary tolerance). No hyperparameter is re-tuned for the new task. The class-balancing weight is recomputed from building label frequency using the same formula applied to shadow frequency. The resulting upper-bound mIoU lands in [0.74, 0.81] across cities, confirming the model is not in a data-starved regime for the task.
Replication outcomes. Pooling the three shadow MAMNet-vanilla cells with the three INRIA MAMNet-vanilla cells gives for the cross-task test, where the smallest -value any sign test can return is 0.016. Three findings transfer cleanly at the cell level: the foreground SP gap, the foreground-vs-background asymmetry, and the cross-entropy under-reporting each go in the same direction on all 6 cells, with cell-level effect sizes Cohen’s and sign-test each. The predicted-positive logit erosion transfers in direction on 5 of 6 cells (Wilcoxon , ). One signature does not transfer: the direction of class-stratified ECE flips between tasks. On INRIA Vienna, ECE rises by 0.138 on predicted-negative pixels and only 0.034 on predicted-positive pixels, the opposite of the shadow pattern.
Interpretation. The class-stratified ECE direction is task-dependent, tracking which class is the population majority (sparse foreground in shadow, dense in INRIA), rather than reflecting a property of geographic LOCO. The other findings depend on properties of the foreground class under shift and replicate across both task semantics. We therefore retain the foreground-asymmetry, under-reporting, logit-erosion, and recoverability claims as properties of geographic LOCO across dense prediction, and explicitly scope the class-stratified ECE result to shadow.
Component-Retraining Experiments
This appendix details the two controlled retraining experiments underlying § 4.4, which localize the recoverable failure component to the decoder for DINOv3, the encoder for OGLANet, and both for MAMNet.
Experiment A (decoder retraining). For each (architecture, holdout city) pair under vanilla LOCO, we freeze the encoder weights, reinitialize the decoder, and retrain only the decoder for 30 epochs on a randomly sampled 25% of the holdout-city training set (≈ 110 images). The encoder receives no gradient updates. The retrained model is evaluated on the held-out city’s 150 test images by per-image foreground IoU with a boundary-tolerant band.
Experiment B (encoder retraining). The inverse setup: freeze the decoder, reset the encoder to its pretrained initialization (ImageNet for the CNN encoders, original DINOv3 weights for the ViT), and retrain only the encoder on the same 25% subset. The decoder receives no gradient updates.
Recovery ratio and exclusion. Recovery is measured by the per-image ratio . To avoid division by zero on images where the within-city gap is negligible, we exclude images with . Pooling per architecture across the three holdout cities, we obtain 335–351 valid values per architecture per experiment. Image-cluster bootstrap () gives the medians and CIs reported in § 4.4; the error bars of Figure 3(c) come from a separate bootstrap () of the same per-image values. Each cell is one training run, and OGLANet values carry the nondeterminism of Appendix § 13.8. The pattern, summarized in Table 40, shows the central architecture-dependent locus claim: each architecture is recovered by one experiment and harmed or unaffected by the other.
| Architecture | Exp. A: decoder | Exp. B: encoder |
| DINOv3 | 0.69 [0.60, 0.74] | −0.13 [−0.27, −0.03] |
| MAMNet | 0.67 [0.63, 0.72] | 0.88 [0.83, 0.90] |
| OGLANet | 0.12 [0.05, 0.21] | 0.86 [0.80, 0.89] |
Table 40. Recovery ratio per architecture, both experiments. Median per-image with image-cluster bootstrap 95% CI. Decoder retraining (Exp. A) recovers DINOv3 and MAMNet but not OGLANet; encoder retraining (Exp. B) recovers MAMNet and OGLANet but actively harms DINOv3.
Why DINOv3’s encoder retraining harms. The negative DINOv3 Exp. B result reflects tight encoder-decoder coupling: the LOCO decoder was trained jointly with a pretrained ViT encoder that barely moved during fine-tuning (whole-encoder relative change of 0.019 from initialization, Appendix § 15.§ 4). Resetting the encoder to its pretrained initialization changes the feature distribution slightly, and the frozen decoder cannot adapt to the new feature space. The performance drops below the LOCO baseline because the decoder is calibrated to a feature space the new encoder no longer produces.
Implication for adaptation method choice. The architecture-dependent locus directly explains the catastrophic failures in Table 2. MRFP+ perturbs encoder features uniformly without targeting OGLANet’s encoder-locus failure, so it costs OGLANet −14.0 mIoU against −4.9 on DINOv3. ISW whitens encoder feature covariance, which corrupts the pretrained representation DINOv3 depends on, so it costs DINOv3 −18.5 mIoU against −11.1 on OGLANet. The architecture-dependent recovery component predicts which adaptation methods will fail catastrophically on a given architecture before the method is run.
Alternative Explanations for the LOCO Gap
This appendix reports the protocol, full results and caveats of the four tests that address alternative explanations of the gap: training-set scale (§H.1), appearance shortcuts (§H.2), fine-tuning of the encoder (§H.3), and the pretraining and fine-tuning regime (§H.4). All four use two-class boundary-tolerant mIoU (±2 px) at high resolution.
Training-set scale
Increasing the LOCO training set from 225 to 900 images narrows the gap by 0.80 mIoU (95% CI [0.46, 1.14]) and leaves the upper bound significantly ahead in every 900-image cell.
Protocol. We train the Vanilla LOCO model and the within-city upper bound of MAMNet, OGLANet and DINOv3 on 225, 450 and 900 training images (rungs), for three held-out cities and three seeds. A held-out city contains 450 high-resolution training images, so at rung 900 the LOCO model trains on 900 images and its upper bound on 450; we report that quantity as the budget-asymmetric gap. Every model selects its checkpoint on 150 validation images at every rung, and every gap is paired per image on the same 150 test images per city. Per-rung and per-architecture intervals resample cells ().
Per-cell gaps and contrasts. Every per-cell gap differs from zero after Holm correction (largest raw ). All are positive except one rung-225 cell, OGLANet/Phoenix seed 1, where LOCO exceeds the upper bound by 2.24 mIoU (). The 225-to-900 contrast is 0.80 mIoU (95% CI [0.46, 1.14]), the 450-to-900 contrast 0.55 mIoU (95% CI [0.32, 0.77]), and the 225-to-450 contrast 0.14 mIoU (95% CI [−0.17, 0.44]) (). These contrasts pool per-image differences of differences across (architecture, city, seed) triples and resample images, so their intervals treat images as independent within a triple.
Collapsed models. We exclude a gap when its LOCO or upper-bound model predicts shadow on more than half of the valid region, pooled over the cell’s images. We set this threshold from the ground truth: shadow occupies at most 0.086 of the valid region in any city, so a model above the threshold has foreground IoU of at most 17.2. The rule excludes three rung-225 gaps (OGLANet/Phoenix seed 2, DINOv3/Miami seed 0, DINOv3/Phoenix seed 0) and none at rungs 450 or 900. With these gaps included, the rung-225 mean is 2.12 ([−4.39, 6.81], 27 cells, 3 sign changes); the other rungs are unchanged.
| Rung 225 | Rung 450 | Rung 900 | |
| MAMNet | 7.31 [6.01, 8.51] (9) | 6.68 [4.92, 8.97] (9) | 6.48 [4.98, 8.18] (9) |
| OGLANet | 7.44 [4.32, 9.81] (8) | 7.31 [5.77, 8.88] (9) | 5.81 [4.74, 7.26] (9) |
| DINOv3 | 4.31 [3.61, 5.30] (7) | 4.15 [2.79, 5.63] (9) | 4.20 [2.38, 6.41] (9) |
| All | 6.48 [5.22, 7.66] (24) | 6.04 [5.00, 7.20] (27) | 5.50 [4.51, 6.53] (27) |
Table 41. Gap by rung. Upper bound minus LOCO, mean over cells with a cell-level bootstrap 95% CI and the number of cells in parentheses. Rung 900 is the budget-asymmetric gap. Collapsed models are excluded (see below).
Appearance alignment. The grayscale histogram matching of § 4.2 improves none of the 27 ladder models it was applied to. One of them is the collapsed DINOv3/Miami rung-225 seed-0 model, whose matched prediction is degenerate. This count carries no interval or p-value.
Scope. The ladder covers high resolution, Vanilla training and budgets up to 900 images, the most our three cities supply.
Training-time appearance augmentation
Protocol. We retrain both the LOCO model and its within-city upper bound with one strong appearance-only augmentation (colour, illumination, contrast, blur and noise; geometry preserved) on the 9 high-resolution (architecture, held-out city) cells at 450 training images, with one to three seeds per side. The gap here is LOCO minus upper bound, so negative values mean LOCO is below its upper bound. Before the grid ran, we fixed the rule: a cell’s gap counts as changed if augmentation moves it by more than twice that cell’s standard deviation of the Vanilla gap over three seeds.
Result. Six of nine gaps do not change (Table 42). The mean gap over the nine cells moves from −6.04 to −4.11 mIoU, and the most negative augmented gap is OGLANet/Chicago at −7.64. On OGLANet/Phoenix the augmented LOCO model exceeds its upper bound by 1.80 mIoU, against a Vanilla gap of −4.50 with a seed SD of 0.24. In the three changed cells, most of the change comes from the LOCO side (LOCO mIoU +3.56, −2.20 and +5.79 on OGLANet/Phoenix, DINOv3/Miami and DINOv3/Phoenix).
| Arch. | City | Vanilla gap | SD | Aug. gap | Threshold | |Δ| | Seeds | Changed |
| MAMNet | Chicago | −4.52 | 0.90 | −3.76 | 1.80 | 0.75 | 1/1 | no |
| MAMNet | Miami | −5.68 | 1.57 | −6.33 | 3.13 | 0.65 | 2/1 | no |
| MAMNet | Phoenix | −9.83 | 4.26 | −5.62 | 8.52 | 4.21 | 3/1 | no |
| OGLANet | Chicago | −9.08 | 2.47 | −7.64 | 4.95 | 1.43 | 3/2 | no |
| OGLANet | Miami | −8.33 | 0.95 | −7.06 | 1.90 | 1.27 | 1/1 | … |
| OGLANet | Phoenix | −4.50 | 0.24 | 1.80 | 0.48 | 6.30 | 3/1 | yes |
| DINOv3 | Chicago | −3.77 | 0.90 | −3.15 | 1.80 | 0.62 | 1/1 | no |
| DINOv3 | Miami | −1.92 | 0.84 | −3.93 | 1.68 | 2.01 | 1/1 | yes |
| DINOv3 | Phoenix | −6.76 | 1.43 | −1.25 | 2.86 | 5.51 | 2/1 | yes |
Table 42. Appearance augmentation, per cell. Gap = LOCO minus upper bound (mIoU points). SD: Vanilla gap standard deviation over three seeds. Threshold = . : absolute change of the gap under augmentation. Seeds: augmented LOCO/upper.
Robustness of the verdicts. Rerunning every model that reached the 100-epoch limit with a 200-epoch limit leaves all nine verdicts unchanged; on DINOv3 the longer run also changes the learning-rate schedule. Replicating OGLANet/Phoenix at three seeds per side, one augmented upper-bound seed collapsed (foreground IoU 1.31); without it the augmented gap is +1.43 and the cell still counts as changed.
Caveats. The rule compares against the Vanilla seed spread and carries no confidence interval, -value or multiplicity correction. Seed counts on the augmented side are one to three and unequal. The test uses one augmentation strength, so an unchanged gap is not attributable to a weak dose. Phoenix has the largest Vanilla seed variance of the three cities and also ranks first on 12 of 20 appearance-distance measures; with three cities, those two explanations of the Phoenix changes cannot be separated. OGLANet cells here are inferred at batch size 1, as are the Vanilla cells they are compared with. CrossEarth was added to the project after this test ran and is not included.
Frozen DINOv3 encoder
Freezing the DINOv3 encoder leaves a gap in all three cities, so fine-tuning the encoder on the source cities is not necessary for the gap. We freeze the DINOv3 encoder (dinov3_vits16) and train only the decoder, for both the LOCO model and the within-city upper bound, on 450 training images at high resolution with one seed.
Gap. The upper-minus-LOCO gap is 7.103 mIoU (95% CI [5.876, 8.348]) in Chicago, 2.669 ([1.729, 3.599]) in Miami and 2.754 ([1.652, 3.939]) in Phoenix, each Holm over the three cities, and the city mean is 4.175 (two-class boundary-tolerant mIoU at , 150 paired test images per city). Images with empty ground-truth foreground (11, 38 and 21 of 150) are retained on both sides of each gap. With one seed, these gaps carry no seed-to-seed interval, and we do not test them against the fine-tuned DINOv3 gap of 4.15 (Appendix § 15.§ 1).
Decoder retraining with a frozen encoder. Retraining the decoder on target labels with the encoder held fixed gives per-city median per-image recovery ratios of −2.674 (Chicago), −7.087 (Miami) and −0.956 (Phoenix) when the decoder is reinitialized; in Chicago and Miami the retrained decoder predicted background only and scored below the LOCO model. Initializing the decoder from the LOCO checkpoint gives , 0.602 and 0.646. With the encoder frozen, the decoder is the only trained component, so recovery through the decoder is expected by construction, and this experiment does not locate the failure; Appendix § 15.§ 4 tests the locus. Both runs differ from Experiment A of § 4.4 in the frozen encoder, and the second also in the decoder initialization, so their values are not comparable with Table 40. Images whose upper-minus-LOCO IoU difference is below 0.02 are excluded, leaving 128, 93 and 107 of 150, and per-image values were not stored, so the medians carry no interval.
Pretraining and fine-tuning regime
Design. DINOv3 differs from our CNNs in self-supervised pretraining and a low fine-tuning learning rate. We separate the two on MANet with a ResNet-50 encoder in a design: ImageNet-1k supervised or DINO self-supervised pretraining, each fine-tuned at learning rate (high) or (low), for three held-out cities and three seeds at 450 training images. We run Experiments A (decoder reinitialized and retrained on held-out labels) and B (encoder retrained) of §4.4 on every cell and report the per-image recovery ratio , pooled over cities and seeds with an image-cluster bootstrap (). Images whose headroom (upper minus LOCO) is below 0.02 are excluded because 's denominator vanishes; this removes 31% to 41% of images per version.
Experiment B. Encoder retraining recovers all four versions: median is 0.82 (95% CI [0.79, 0.86]) for supervised/high, 0.85 ([0.81, 0.88]) for supervised/low, 0.87 ([0.84, 0.90]) for self-supervised/high and 0.80 ([0.76, 0.83]) for self-supervised/low. The self-supervised/low version recovers less than supervised/high (mean per-image difference −0.13, [−0.22, −0.05], Holm ) and than self-supervised/high (−0.16, [−0.24, −0.08], Holm ), and is level with supervised/low (−0.04, [−0.10, +0.04], ); Holm is over these three contrasts. The main effects of pretraining (−0.003) and learning rate (−0.018) on median are small.
Experiment A. Decoder retraining gives median of 0.64 ([0.59, 0.67]) for supervised/high, −0.14 ([−0.24, −0.03]) for supervised/low, 0.58 ([0.54, 0.60]) for self-supervised/high and 0.12 ([0.03, 0.18]) for self-supervised/low. The two low-learning-rate versions recover less through the decoder. In this design the retrained decoder uses the cell’s own fine-tuning learning rate, and 29 of the 36 Experiment A cells reach the 30-epoch budget, so the low Experiment A values may reflect a slower decoder rather than the regime.
Encoder drift. The whole-encoder relative change from initialization has median 0.017 (supervised/low) and 0.032 (self-supervised/low), against 0.019 for the fine-tuned DINOv3 LOCO encoders, and 0.22 and 0.38 for the two high-learning-rate versions. Drift tracks the learning rate rather than the pretraining objective.
Caveats. The design varies the pretraining objective at fixed data: DINO-ResNet-50 is pretrained on ImageNet-1k (1.28M images) and DINOv3 on LVD-1689M, and no public self-supervised CNN closes that gap, so pretraining scale remains confounded. Experiments A and B use held-out-city labels and are diagnostic, not a deployable adaptation. The design covers one architecture.
External Benchmark: S-EO
The gap appears on S-EO [4], an external benchmark that differs from ShadowTransfer in sensor (WorldView-3), label source (shadows derived from a digital surface model, DSM) and cities.
Protocol. The three S-EO cities are Omaha, Jacksonville and San Diego. For each held-out city, the LOCO model trains on 3,500 patches from the two other cities and the within-city upper bound on 3,500 patches from the held-out city. We train MAMNet, OGLANet, DINOv3 and CrossEarth with three seeds each and evaluate on 1,637 patches from 42 areas of interest (AOIs) in Omaha, 2,642 from 72 in Jacksonville and 1,870 from 52 in San Diego. The metric is two-class boundary-tolerant mIoU at with pixels the DSM did not observe excluded. Intervals are AOI cluster bootstrap intervals () with seed as a stratum, and -values come from AOI-level sign-flip permutation tests with Holm correction over the 12 (architecture, city) pairs. OGLANet runs at inference batch size 1, because its scores depend on the inference batch size in our reimplementation.
Gap. The upper-minus-LOCO gap is positive in all 12 (architecture, city) pairs (Holm each), from 2.10 mIoU (OGLANet, Omaha) to 7.95 (CrossEarth, Omaha). Pooled within architecture, the gap is 6.98 (95% CI [6.50, 7.46]) for MAMNet, 6.93 ([6.44, 7.42]) for CrossEarth, 4.28 ([3.97, 4.59]) for DINOv3 and 3.45 ([3.19, 3.71]) for OGLANet, and 5.41 ([5.07, 5.75], 664 AOI units) over all four. We do not pool or test the S-EO gaps against the ShadowTransfer gaps, because the two benchmarks differ in sensor, labels, training budget and test unit.
Diagnostics of Section 4. We repeat three diagnostics on S-EO. (i) Mean predicted shadow probability on ground-truth shadow, , is lower under LOCO than under the upper bound in 10 of 12 pairs (Holm ) and higher in none, with seed-averaged differences from (MAMNet, San Diego) to (OGLANet, Omaha). The background control also differs in 10 of 12 pairs, so on S-EO the confidence shift is not confined to the shadow class. The foreground change is smallest in Omaha, a mean absolute of 0.0161 against 0.0889 in the other two cities. (ii) The foreground-minus-background AURC gap (LOCO minus upper bound; positive means higher selective risk under LOCO) is positive in 12 of 12 pairs and Holm-significant in 8, pooling to ([+0.0422, +0.0539], 660 AOI units). (iii) The sign of the decoder-retraining recovery (Experiment A) separates by architecture in the same direction as on ShadowTransfer: positive in 18 of 18 MAMNet and DINOv3 cells and negative in 9 of 9 OGLANet cells, under both foreground-only and two-class differencing. A cell’s gap predicts the magnitude of its recovery (), so gap size is an alternative account of recovery magnitude; the sign separation holds where the DINOv3 and OGLANet gaps overlap. CrossEarth has no Experiment A implementation.
Appearance and sun geometry. Grayscale histogram matching of target to source improves 0 of 36 S-EO cells. Regressing the per-patch gap on the patch’s sun-elevation distance to the LOCO training pool, with city and architecture fixed effects, gives a pooled slope of with under city-clustered standard errors (3 clusters) and under ordinary least squares, which ignores the clustering. The data give no support to a sun-geometry account, and with three clusters the test has little power.
Adaptation methods. Of 27 (architecture, method) pairs, 21 score below Vanilla LOCO and 5 are mixed. MAMNet with MRFP+ fails to train in all nine of its cells, and no method cell collapses at test time.
SIB. At seed 0, SIB’s worst (architecture, city) cell is OGLANet/San Diego at 61.15 mIoU, 1.57 below Vanilla’s worst cell, OGLANet/Omaha at 62.72. SIB scores significantly below Vanilla in 7 of 12 cells and above it in 1 (Holm over 12).
Decomposition of SIB’s S-EO deficit. On a single-seed, two-architecture subset of the S-EO grid, the DINOv3 SIB configuration with FDA sits 1.10 mIoU below Vanilla on average over the three cities, and the SIB core without FDA sits 0.66 below. On MAMNet, adding skip attention gates (SAG) to the core changes the score by +0.28. These values suggest that part of the deficit comes with the FDA component; the grid was not designed as an ablation, and we did not test this.
Caveats. S-EO labels are generated from a DSM; both sides of every gap use the same labels on the same patches, so label noise affects absolute levels more than the gap. 108 test patches (29 in Omaha, 79 in San Diego) are entirely DSM-unseen, score 0 in every cell of their city, and contribute a difference of exactly 0 to every gap. Five cells stopped at the epoch ceiling before early stopping triggered. 53 of the 72 Vanilla cells carry no inference-noise band, so no AURC value has a run-to-run interval.
SIB Implementation Details
SIB sits between the encoder and the decoder of any backbone. Encoder features are decomposed via a single-level 2D Haar wavelet into four subbands , each in . The Haar transform uses no learned parameters and is perfectly invertible; gradients propagate cleanly through the forward and inverse operations. Compressed subbands are recombined by inverse Haar before being passed to the decoder.
Content VIB on . The content subband carries surface appearance and city-specific texture. We compress it through a per-pixel variational information bottleneck. Two convolutional MLPs map to per-pixel mean and log-variance . At training time we sample with ; at inference we use the mean only. The KL penalty against the standard Gaussian prior is weighted by an intensity-conditioned coefficient
where is the mean pixel intensity in a local window around spatial location , computed from the original RGB input and downsampled to the bottleneck resolution. The intensity MLP is a small stack with ReLU.
Edge VIB on and . The horizontal and vertical boundary subbands carry shadow geometry. They share a single Edge VIB module with reduced channel width ( hidden) and a small fixed weight ( on MAMNet). The diagonal subband passes through unmodified.
Skip Attention Gates. For U-Net architectures (MAMNet, OGLANet), each skip connection passes through a channel-wise sigmoid gate computed by global average pooling followed by a linear layer. Gate biases initialize at +2.0 so that sigmoid output starts near 0.88, making the gate a near-identity at initialization. The model learns to suppress city-specific channels routed around the bottleneck. SAG is not applied for DINOv3.
CNN-only preprocessing. For CNN architectures, we apply Fourier Domain Adaptation as input-level preprocessing with attenuated spectral radius , and append a contrast channel as a fourth input channel computed as the local intensity range over a window. ViT inputs receive neither addition.
Coverage thresholding at deployment. For each cell, we fit on the source-city validation split as the largest coverage at which selective shadow error satisfies , with globally. The fitted is then applied at inference on the held-out city. The coverage grid is . No held-out labels enter the procedure.
Training loss. The full training objective combines the boundary-tolerant segmentation loss with the two band-specific KL terms:
The KL terms ramp from 0 to their target weight linearly over the first 10% of training iterations to let the encoder-decoder pair settle before compression activates.
Hyperparameters. Table 43 lists all hyperparameters. The CNN-specific row is applied for MAMNet and OGLANet only; the ViT row applies to DINOv3. The reported SIB runs switch off the intensity adaptation of and the content augmentation; Appendix § 19 tests both against a reference that enables them.
| Group | Parameter | Value |
| Content VIB | (constant) | (MAMNet, OGLANet); (DINOv3) |
| VIB hidden width | ||
| Edge VIB | ( on MAMNet) | |
| VIB hidden width | ||
| SAG (U-Net) | gate bias init | +2.0 |
| CNN preproc | FDA spectral radius | 0.005 |
| contrast window | ||
| Deployment | (coverage criterion) | 0.5 |
| coverage grid step | 0.05 | |
| KL warmup | fraction of total iters | 0.10 |
| Optimizer | AdamW, weight decay | 0.01 (OGLANet: Adamax, ) |
| LR | MAMNet | |
| OGLANet | ||
| DINOv3 | ||
| Schedule | cosine, warmup epochs | 5 (OGLANet: reduce on plateau) |
| Batch size | per GPU | 8 |
| Epochs | max (early stop on val mIoU) | 100 |
Table 43. SIB hyperparameters of the reported runs. Values are fixed across all 9 evaluation cells; per-architecture values are given where they differ.
Method Comparison by Gap to the Upper Bound
Table 44 compares every method by its gap to the within-city upper bound over the nine cells of § 5.2.
| Method | Access | Mean [95% CI] | Worst (cell) | Harm/benefit | Largest harm | SD | SIB−method [95% CI] | Holm | |
| SIB | S/U | 9 | 6.42 [5.01, 7.96] | 10.41 (OGLANet/Miami) | 2/2 | −3.79 | 2.45 | — | — |
| FDA | U | 9 | 6.47 [5.34, 7.54] | 9.10 (DINOv3/Phx) | 1/4 | −6.51 | 1.83 | +1.31 [−1.01, +3.65] | 0.227 |
| SegDeSic | U | 9 | 6.84 [4.25, 9.33] | 13.20 (OGLANet/Mia) | 2/3 | −5.49 | 4.09 | −2.80 [−5.16, −0.13] | 0.141 |
| SAR | D | 9 | 6.86 [4.81, 9.05] | 13.02 (DINOv3/Chi) | 2/1 | −0.01 | 3.52 | −2.61 [−4.18, −0.16] | 0.141 |
| Vanilla | D | 9 | 6.91 [4.83, 9.11] | 13.01 (DINOv3/Chi) | 1/1 | −0.01 | 3.55 | −2.61 [−4.18, −0.16] | 0.141 |
| BSM | S | 8 | 8.05 [5.14, 11.46] | 17.40 (OGLANet/Chi) | 6/2 | −6.74 | 4.99 | −6.99 [−9.01, −1.04] | ≤ 0.001 |
| FADA | S | 9 | 9.05 [6.98, 11.17] | 13.84 (MAMNet/Dia) | 5/1 | −9.39 | 3.44 | −3.43 [−6.17, −1.24] | 0.016 |
| IIM | S | 9 | 9.75 [6.37, 14.01] | 24.06 (MAMNet/Phx) | 4/1 | −14.54 | 6.19 | −13.66 [−16.38, −10.19] | ≤ 0.001 |
| MRF+ | S | 9 | 10.11 [6.57, 13.75] | 18.59 (MAMNet/Phx) | 7/1 | −9.07 | 5.77 | −7.18 [−10.86, −5.52] | ≤ 0.001 |
| ISW | S | 9 | 13.83 [10.22, 17.69] | 23.28 (MAMNet/Phx) | 7/0 | −13.76 | 6.16 | −12.87 [−15.54, −10.45] | ≤ 0.001 |
Table 44. Gap to the within-city upper bound for every method. Nine cells (MAMNet, OGLANet, DINOv3 × Chicago, Miami, Phoenix), high resolution, two-class boundary-tolerant mIoU (±2 px) in percentage points, rows ordered by mean gap. Access: S source-only, U unlabeled target imagery, D deployment-time. Mean: mean gap with a cell-level bootstrap 95% CI (B = 10,000). Worst: largest gap, with its cell. Harm/benefit: cells significantly worse or better than Vanilla (paired image bootstrap, Holm over the method’s cells). Largest harm: smallest method-minus-Vanilla cell mean. SD: spread of the per-cell gap. SIB−method: SIB’s worst-case gap minus the method’s, with an image-cluster bootstrap 95% CI and Holm-adjusted over the nine comparisons; negative values favour SIB.
Relative gaps. As a percentage of the upper bound, SIB’s mean gap is 8.52% and FDA’s 8.53%; their worst-case relative gaps are 14.08% and 13.42%.
SIB against FDA per cell. SIB is significantly closer to the upper bound than FDA in 2 of 9 cells, farther in 2, and not separable in 5 (Holm over 9).
Cells that reach the upper bound. Every method’s gap is significant in every cell except SegDeSiC on DINOv3/Phoenix, the one cell whose method score exceeds its upper bound.
Scope and caveats. We chose this profile after the pre-committed comparison of worst-case gaps had been read, so it is exploratory. The grid is the one behind Table 2: the 27 cells that originally trained on 900 images (MAMNet Vanilla and its six adaptation methods, OGLANet SegDeSiC and OGLANet SIB) are 450-image retrains, and every other cell is the published run. Each cell is one training seed, except MAMNet-SIB, whose value is the mean of three seeds; the bootstraps resample test images only. The three OGLANet-SIB cells reached the 100-epoch limit with validation still improving, so they are lower bounds on the recipe; their FDA donor pool is the held-out city’s 450 unlabeled training images. OGLANet inference depends on batch size and random state through a defect in our DFFM reimplementation (Appendix § 13.8; up to 0.66 mIoU across seeds on an OGLANet-SIB checkpoint), and every OGLANet value is a single pass. BSM has 8 cells because its DINOv3/Chicago model collapsed. SAR rows are Vanilla plus SAR’s within-pass change, so SAR’s harm/benefit counts test only whether that change differs from zero. The worst-case difference is a difference of two maxima and has no unbiased bootstrap, so its interval is descriptive. Harm, benefit and significance counts depend on and the stated Holm families.
Extended SIB Component Ablations
This appendix extends Table 3 with every ablation run during SIB development, 10 rows per CNN architecture and 8 for DINOv3. Table 3 reports the five rows that isolate architecture-specific gating (SAG), CNN input preprocessing (FDA), spectral decomposition (Haar), content compression ( VIB) and subband targeting (the inverse-evidence row).
Ablation methodology. Each ablation switches off (or, for the inverse-evidence row, replaces) one design choice of a full-model reference trained in the same batch. Each is the ablation’s two-class boundary-tolerant mIoU minus the reference’s, per image, pooled over the three held-out cities at high resolution ( paired images). The interval and come from a paired image bootstrap that resamples 150 images within each city (, two-tailed, with replicates past zero). One bootstrap draw is shared by all tests, so intervals of different rows are correlated. Holm correction runs over all 28 (row, architecture) tests. 10 of the 28 tests sit at the bootstrap floor , so their is resolution-limited.
Reference and training differences. The batch reference differs from the published SIB runs: the reference enables intensity-adaptive and content augmentation, which the published runs switch off, and on MAMNet and OGLANet the reference uses early-stopping patience 15 where the published runs and every ablation use 25. Each CNN therefore also carries this patience change. The published configuration minus the reference ranges from −1.12 to +2.80 mIoU per (architecture, city) and changes sign across cities within OGLANet and DINOv3. Three rows switch more than one flag. The—FDA rows also disable the contrast stage. The MAMNet—All VIB row disables intensity-adaptive . MAMNet has no run that removes only the VIB, and its SIB has no cross-source mixing.
Training and scoring conditions. Each cell is one training run with no recorded seed, and the intervals resample test images only. The OGLANet ablations and their reference trained on a 900/300/150 train/validation/test pool, not the 450-image protocol of the main tables. The MAMNet and DINOv3 runs record no training count. OGLANet predictions come from a single fp16 inference pass at batch size 8 and carry the nondeterminism of Appendix § 13.8. The MAMNet and OGLANet runs name the held-out city as FDA donor root. OGLANet’s trainer does not pass that root to the dataset, and we did not establish whether the MAMNet runs used held-out-city images as donors.
Cells are scored from the stored prediction masks. The checkpoints are no longer available, so no cell could be re-inferred, and no probability-based metric is reported.
| Variant | MAMNet | OGLANet | DINOv3 |
| Reference (mIoU) | 73.60 [72.17, 75.03] | 67.79 [66.66, 68.93] | 70.13 [68.91, 71.33] |
| − VIB on | – | −1.01** [−1.52, −0.50] | −0.71 [−1.30, −0.19] |
| − All VIB | +0.29 [−0.33, +0.94] | – | – |
| − Edge VIB | +0.04 [−0.49, +0.60] | – | – |
| Uniform | +0.03 [−0.61, +0.70] | −1.57** [−2.11, −1.01] | −0.26 [−0.72, +0.22] |
| Symmetric VIB | −0.76 [−1.37, −0.17] | −0.56 [−1.09, −0.02] | −0.27 [−0.78, +0.23] |
| − Content aug. | +0.30 [−0.27, +0.87] | −1.68** [−2.21, −1.15] | −0.94** [−1.37, −0.51] |
| − Cross-source mixing | N/A | −2.18** [−2.80, −1.57] | −0.48 [−0.95, −0.03] |
| Aug. on all subbands | +0.16 [−0.38, +0.74] | −1.43** [−2.03, −0.81] | −0.01 [−0.46, +0.44] |
| − SAG | +0.23 [−0.40, +0.87] | −2.06** [−2.58, −1.55] | N/A |
| − FDA, contrast | −4.42** [−5.37, −3.49] | −0.87 [−1.51, −0.23] | N/A |
| − Haar | +0.11 [−0.59, +0.79] | −0.63 [−1.18, −0.08] | −23.49**‡ [−24.83, −22.09] |
| VIB on (inv.) | −0.29 [−0.84, +0.27] | −1.99** [−2.53, −1.45] | −0.74 [−1.25, −0.24] |
Table 45. Full SIB component ablations. Mean mIoU against each architecture’s batch reference with 95% CI, pooled over the three held-out cities ( paired images). Stars: Holm-adjusted over 28 tests, * , ** . N/A: component absent from that architecture’s SIB. –: not run. Uniform : no intensity adaptation. Symmetric VIB: the same on , and . Edge VIB: the VIB on . −Haar: one VIB on the undecomposed . Inv.: VIB on in place of . ‡Collapsed run (text).
Reading the table. OGLANet shows the most significant rows: removing the content VIB, uniform , removing content augmentation, removing cross-source mixing, augmenting all subbands, removing SAG and compressing in place of . On MAMNet only the −FDA row is significant. On DINOv3 removing content augmentation and removing the Haar decomposition are significant. The DINOv3 −Haar runs predict almost no shadow. On Chicago and Miami the predicted-positive fraction is 0.0000 against ground-truth fractions of 0.0864 and 0.0802, 100% and 99% of test images contain no predicted shadow, and foreground IoU is 0.0 against the reference’s 56.1 and 49.3. On Phoenix the fraction is 0.0026 against 0.0347, and foreground IoU is 0.2 against 25.8. This row therefore compares a collapsed training run with the reference.
Coverage Threshold Robustness
The coverage thresholding recipe of § 5.1 is governed by a single global hyperparameter that sets the source-val error reduction criterion. The main-text reports use (halve the source-val full-coverage error). This appendix verifies that the result is stable under perturbation of .
Source-val operating-point verification. Before testing on the held-out city, we confirm that the chosen achieves the -criterion on the source-val split for every cell. Table 46 reports per-cell , the source-val full-coverage error, the source-val error at , and the achieved reduction percentage. All 9 cells achieve at least the target ( reduction); reductions cluster slightly above 50% (50.1% to 60.6%) because lands on the smallest grid value crossing the criterion. No cell falls back to the trivial .
| Arch | Holdout | reduction | |||
| MAMNet | Phoenix | 0.81 | 0.239 | 0.094 | 60.6% |
| MAMNet | Miami | 0.67 | 0.362 | 0.165 | 54.5% |
| MAMNet | Chicago | 0.72 | 0.334 | 0.167 | 50.1% |
| OGLANet | Phoenix | 0.76 | 0.315 | 0.134 | 57.6% |
| OGLANet | Miami | 0.62 | 0.453 | 0.204 | 55.0% |
| OGLANet | Chicago | 0.67 | 0.399 | 0.178 | 55.5% |
| DINOv3 | Phoenix | 0.72 | 0.373 | 0.163 | 56.4% |
| DINOv3 | Miami | 0.57 | 0.518 | 0.257 | 50.5% |
| DINOv3 | Chicago | 0.62 | 0.456 | 0.221 | 51.5% |
Table 46. Source-val operating-point verification at . For each cell: fitted , source-val full-coverage error, source-val error at , and achieved reduction.
sweep on the held-out city. We re-fit at on the source-val split and apply each fitted operating point to the held-out city. Table 47 reports the population-level test at each . All three settings produce 9/9 cells improving with Wilcoxon (the floor under one-sided rank-sum), with the mean held-out reduction scaling monotonically with the criterion’s aggressiveness. No cell falls back to the trivial coverage at any .
| Mean | Wilcoxon | Cells improving | Median | |
| 0.3 | −0.121 | 0.00195 | 9/9 | 0.81 |
| 0.5 | −0.199 | 0.00195 | 9/9 | 0.67 |
| 0.7 | −0.264 | 0.00195 | 9/9 | 0.57 |
Table 47. Population-level coverage recovery across . For each : mean held-out selective shadow error reduction across 9 cells, one-sided Wilcoxon -value (alternative: ), cells improving, and median .
Per-cell behavior across . Table 48 reports per-cell and held-out selective shadow error reduction for each . The pattern is consistent across cells: smaller yields a less aggressive operating point (larger , smaller magnitude held-out reduction); larger yields a more aggressive operating point (smaller , larger magnitude reduction). The DINOv3-Phoenix cell is the smallest in magnitude across all three settings, consistent with this cell’s small within-city to LOCO gap (per Appendix § 18); even on this cell, all settings produce a non-trivial reduction.
| Arch | Holdout | |||
| MAMNet | Phoenix | −0.046 (0.91) | −0.099 (0.81) | −0.128 (0.76) |
| MAMNet | Miami | −0.148 (0.81) | −0.230 (0.67) | −0.264 (0.38) |
| MAMNet | Chicago | −0.115 (0.86) | −0.212 (0.72) | −0.271 (0.24) |
| OGLANet | Phoenix | −0.085 (0.86) | −0.131 (0.76) | −0.171 (0.67) |
| OGLANet | Miami | −0.119 (0.76) | −0.228 (0.62) | −0.370 (0.48) |
| OGLANet | Chicago | −0.111 (0.81) | −0.226 (0.67) | −0.361 (0.53) |
| DINOv3 | Phoenix | −0.050 (0.81) | −0.085 (0.72) | −0.156 (0.57) |
| DINOv3 | Miami | −0.225 (0.72) | −0.312 (0.57) | −0.349 (0.34) |
| DINOv3 | Chicago | −0.187 (0.76) | −0.268 (0.62) | −0.301 (0.29) |
Table 48. Per-cell coverage recovery across . Held-out selective shadow error at minus error at , with in parentheses.
mIoU at threshold 0.5 is preserved by construction. Coverage thresholding does not modify the 0/1 decision boundary at ; it abstains on low-confidence pixels. The mIoU at threshold 0.5 reported in Appendix § 18 is therefore identical for SIB at and at any in the tested range. The sweep characterizes the magnitude of selective-error recovery; deployment selects as a single fixed global hyperparameter.
References
- [1]Vishal Batchu, Alex Wilson, Betty Peng, Carl Elkin, Umangi Jain, Christopher Van Arsdale, Ross Goroshin, and Varun Gulshan. Satellite sunroof: High-res digital surface models and road segmentation for global solar mapping. In NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning, 2024.arxiv.org/abs/2408.14400
- [2]Longchao Da, Xiangrui Liu, Mithun Shivakoti, T Pranav Kutrallngam, Yezhou Yang, and Hua Wei. Shadebench: A benchmark dataset and method for shade simulation in sustainable society. In NeurIPS 2025 UrbanAI: Harnessing Artificial Intelligence for Smart Cities, 2025.
- [3]S. V. Ukkusuri, S. U. Park, S. Mittal, L. Chapman, G. Manoli, A. Santos, et al. We need to prepare our transport systems for heatwaves—here’s how. Nature, 632(8024):253–256, 2024.DOI
- [4]Elías Masquil, Roger Marí, Thibaud Ehret, Enric Meinhardt-Llopis, Pablo Musé, and Gabriele Facciolo. S-eo: A large-scale dataset for geometry-aware shadow detection in remote sensing applications. In Proceedings of the Computer Vision and Pattern Recognition, EarthVision: Large Scale Computer Vision for Remote Sensing Imagery, pages 2383–2393, 2025.arxiv.org/abs/2504.06920
- [5]S. Mittal and S. V. Ukkusuri. Overcoming data scarcity in transit planning: A novel framework combining machine learning and metaheuristics. Data Science for Transportation, 7(1):4, 2025.DOI
- [6]R. Verma, S. Mittal, and S. V. Ukkusuri. Spatial Access of America: Multiple indicators of accessibility to opportunities. Scientific Data, 12(1):1223, 2025.DOI
- [7]Jean Ogier du Terrail, Samy-Safwan Ayed, Edwige Cyffers, Felix Grimberg, Chaoyang He, Regis Loeb, Paul Mangold, Tanguy Marchand, Othmane Marfoq, Erum Mushtaq, et al. Flamby: Datasets and benchmarks for cross-silo federated learning in realistic healthcare settings. Advances in Neural Information Processing Systems, 35:5315–5334, 2022.arxiv.org/abs/2210.04620
- [8]Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021.arxiv.org/abs/2110.08733
- [9]Ziyang Gong, Zhixiang Wei, Di Wang, Xiaoxing Hu, Xianzheng Ma, Hongruixuan Chen, Yuru Jia, Yupeng Deng, Zhenming Ji, Xiangwei Zhu, et al. Crossaerth: Geospatial vision foundation model for domain generalizable remote sensing semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.DOI
- [10]Shuang Luo, Huifang Li, and Huanfeng Shen. Deeply supervised convolutional neural network for shadow detection based on a novel aerial shadow imagery dataset. ISPRS Journal of Photogrammetry and Remote Sensing, 167:443–457, 2020.DOI
- [11]Tomás F Yago Vicente, Le Hou, Chen-Ping Yu, Minh Hoai, and Dimitris Samaras. Large-scale training of shadow detectors with noisily-annotated shadow examples. In European conference on computer vision, pages 816–832. Springer, 2016.DOI
- [12]Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788–1797, 2018.arxiv.org/abs/1712.02478
- [13]Seyed Majid Azimi and Reza Bahmanyar. Airborne-shadow: towards fine-grained shadow detection in aerial imagery. In DAGM German conference on pattern recognition, pages 34–49. Springer, 2023.DOI
- [14]Lei Zhang, Qing Zhang, Yu Wu, Yanfeng Zhang, Shan Xiang, Donghai Xie, and Zeyu Wang. A full-scale shadow detection network based on multiple attention mechanisms for remote-sensing images. Remote Sensing, 16(24):4789, 2024.DOI
- [15]Xueli Chang, Haiyang Shi, Tiejun Zhang, Huazhong Jin, and Ao Xu. Shadow detection in remote sensing images based on multibranch feature aggregation and channel-spatial attention. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 18:2618–2630, 2024.
- [16]Jiayi Li and Qunming Wang. Csdtformer: A cloud and shadow detection method for landsat images based on transformer. International Journal of Applied Earth Observation and Geoinformation, 129:103799, 2024.DOI
- [17]Yakun Xie, Dejun Feng, Hongyu Chen, Ziyang Liao, Jun Zhu, Chuangnong Li, and Sung Wook Baik. An omni-scale global–local aware network for shadow extraction in remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 193:29–44, 2022. 3, 5, 28DOI
- [18]Xiaowei Hu, Tianyu Wang, Chi-Wing Fu, Yitong Jiang, Qiong Wang, and Pheng-Ann Heng. Revisiting shadow detection: A new benchmark dataset for complex world. IEEE Transactions on Image Processing, 30:1925–1934, 2021. 3DOI
- [19]Xiaowei Hu, Zhenghao Xing, Tianyu Wang, Chi-Wing Fu, and Pheng-Ann Heng. Unveiling deep shadows: A survey and benchmark on image and video shadow detection, removal, and generation in the deep learning era. International Journal of Computer Vision, 134(4):158, 2026. 3, 16arxiv.org/abs/2409.02108
- [20]Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085–4095, 2020. 3, 5, 9, 25
- [21]Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, and Shuaicheng Liu. You only look around: learning illumination-invariant feature for low-light object detection. Advances in neural information processing systems, 37:87136–87158, 2024. 3, 5, 26arxiv.org/abs/2410.18398
- [22]Muying Luo, Shunping Ji, and Shiqing Wei. A diverse large-scale building dataset and a novel plug-and-play domain generalization method for building extraction. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:4122–4138, 2023. 3, 5, 35arxiv.org/abs/2208.10004
- [23]Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11580–11590, 2021. 3, 5, 26arxiv.org/abs/2103.15597
- [24]Sachin Verma, Frank Lindseth, and Gabriel Kiss. Segsdesignet: Lightweight semantic segmentation in remote sensing with geo-coordinate embeddings for domain adaptation. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9093–9104. IEEE, 2025. 3, 5, 25arxiv.org/abs/2503.08290
- [25]Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5904–5914, 2024. 3, 5, 27arxiv.org/abs/2311.18331
- [26]Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. Advances in Neural Information Processing Systems, 37:94047–94072, 2024. 3, 5, 9, 27DOI
- [27]Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 3arxiv.org/abs/2006.10726
- [28]Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofe Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 3, 5, 36arxiv.org/abs/2302.12400
- [29]Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025. 5arxiv.org/abs/2508.10104
- [30]Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. Advances in neural information processing systems, 30, 2017. 7arxiv.org/abs/1705.08500
- [31]Jeremias Traub, Till J Bungert, Carsten T Lüth, Michael Baumgartner, Klaus H Maier-Hein, Lena Maier-Hein, and Paul F Jäger. Overcoming common flaws in the evaluation of selective classification systems. Advances in Neural Information Processing Systems, 37:2323–2347, 2024.arxiv.org/abs/2407.01032
- [32]Luís Felipe P Cattelan and Danilo Silva. How to fix a broken confidence estimator: Evaluating post-hoc methods for selective classification with deep neural networks. arXiv preprint arXiv:2305.15508, 2023. 7arxiv.org/abs/2305.15508
- [33]Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez. Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark. In 2017 IEEE International geoscience and remote sensing symposium (IGARSS), pages 3226–3229. IEEE, 2017. 7, 44DOI
- [34]Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 9arxiv.org/abs/1612.00410
- [35]Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. ACCESS: Advancing innovation: NSF’s advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing (PEARC ’23). ACM, 2023. 11
- [36]Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86–92, 2021. 16arxiv.org/abs/1803.09010
- [37]CVAT.ai Corporation. Computer vision annotation tool (cvat). 16, 17
- [38]Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023. 16, 17arxiv.org/abs/2304.02643
- [39]Shagun Mittal, Raymond A. Yeh, and Satish Ukkusuri. ShadowTransfer (revision 5bef90e), 2026. 17
- [40]J Richard Landis and Gary G Koch. The measurement of observer agreement for categorical data. biometrics, pages 159–174, 1977. 19pubmed.ncbi.nlm.nih.gov/843571
- [41]Teja Kattenborn, Felix Schiefer, Julian Frey, Hannes Feilhauer, Miguel D Mahecha, and Carsten F Dormann. Spatially autocorrelated training and validation samples inflate performance assessment of convolutional neural networks. ISPRS Open Journal of Photogrammetry and Remote Sensing, 5:100018, 2022. 22DOI
- [42]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 24arxiv.org/abs/1512.03385
- [43]Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldien El-Nouby, et al. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 35arxiv.org/abs/2304.07193
- [44]Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger, fewer, & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28619–28630, 2024. 35arxiv.org/abs/2312.04265
- [45]Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1280–1289, 2022. 35arxiv.org/abs/2112.01527
- [46]Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1510–1519, 2017. 35arxiv.org/abs/1703.06868
- [47]Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 36arxiv.org/abs/2010.01412