WorksComputer Science
Forecasting Earth Surface Change in Embedding Space
Abstract
Repeated satellite imagery permits understanding of how our planet changes over time, letting us predict its future and plan mitigating actions. Most existing studies that rely on a time series of satellite observations forecast the spectral values at each pixel for a future target date. However, clouds, shadows, and sensor noise often corrupt these observations, so that spectral forecasters waste much of their capacity on low-level features or noise rather than real change. With the advent of Earth foundation models, we argue that forecasting in embedding space is a feasible and more effective alternative. To this end, we design a token space that handles irregular histories from different sensors and foundation models and supports forecasts at any query date. Tokens from this space are inputs to our proposed forecaster, EForecast, which is a transformer with two branches: a Gaussian branch that directly predicts a mean point forecast and per-dimension standard deviations, and a rectified-flow branch that generates forecast ensembles. On Sentinel-1 and Sentinel-2 benchmarks, forecasting embeddings instead of spectral values halves the cosine distance (from 0.20 to 0.10) and greatly reduces overconfidence, lowering the z-score standard deviation from 2.63 to 1.22. By comparing new observations with its forecasts, EForecast can also detect anomalous changes without labels or retraining, raising F from 41.58 to 48.70 on a global benchmark. This consistent superiority is validated on nine remote sensing foundation models. Beyond change detection, we believe our approach advances the field by permitting effective reuse of diverse pretrained task heads. Our code is publicly available at https://github.com/ucam-eo/e2forecast.
Introduction
The Earth’s surface is dynamic, where crops grow and are harvested, forests burn and regenerate, cities expand, and populations move. Most of these changes follow complex, seasonal, and long-term patterns. Understanding these patterns allows us to forecast the future and prepare in advance, which is valuable for food security, disaster response, and environmental management.
The remote sensing community has become increasingly proficient in forecasting future satellite images [23, 53, 74] and physical variables [4]. However, most studies model both the dynamics and the forecast in the original signal (spectral) space. Consequently, the model has to deal with artifacts such as clouds, shadows, and sensor noise, which vary between acquisitions but say little about how the land changes [59]. We demonstrate § 1.1 that signal-space forecasts lead models to miss many high-level features. Moreover, due to significant differences in sensor modalities, such as SAR and optical, it is difficult to forecast flexible multi-modal observations.
In recent years, geospatial foundation models have been used to create embeddings that turn noisy and heterogeneous observations into compact and semantically meaningful vectors [7, 17, 18, 24, 61]. These embeddings suppress much of the acquisition-specific variation while preserving the semantic content of the observations. Their wide impact can be witnessed in diverse downstream tasks such as crop mapping, landslide detection, and habitat analysis [11, 46, 54]. Existing work uses embeddings to describe what has already been observed, but does not explicitly model their evolution to forecast future states. Instead, we seek to directly forecast in the embedding space and exploit different foundation models that can generate embeddings for different sensors, spatial resolution, patch size, and time durations.
Unlike image-space forecasting, which predicts a small number of spectral channels per pixel, embedding-space forecasting predicts a high-dimensional feature vector for each token. This poses two main challenges: (i) How can we design a token space that supports forecasts from irregularly timed inputs? Satellite observations are rarely regular and revisit intervals differ between sensors and clouds can remove blocks of observations, so two consecutive embeddings may be days or months apart. Most forecasting models assume input observations are spaced by regular steps and predict assuming a fixed lead time [41, 44]. (ii) How can we model multiple possible outputs? Even with a complete embedding history, the future, by definition, remains uncertain. For example, the croplands can be either rotated or left fallow, or a seashore may or may not have vessels. A forecast should ideally generate an ensemble of potential futures, each with a corresponding uncertainty. Existing models (§ 2) only partly meet this requirement. For example, probability forecasting models give the most likely future distribution and its uncertainty, but it is difficult to draw realistic, potentially correlated features from high-dimensional distributions [39, 55]. Some generative models can produce realistic ensemble forecasts [15, 33, 38], but their most likely future and uncertainty are only known after drawing at least 100 samples with tens of denoising steps, which makes it inefficient.

Figure 1. A demo of EForecast’s flexible forecasting with any query date, producing both point and ensemble forecasts with uncertainty. We choose this Cambridge scene because it includes both urban and rural areas. With time-series OlmoEarth embeddings, we train EForecast on 2024 only and use January 2025 as context, so every later forecast comes from learned temporal patterns rather than copied observations. Both point and ensemble forecasts of 10 generated embeddings capture the seasonal changes effectively, while ensemble forecasts can better describe the diversity and details.
Forecasting in embedding space
We propose EForecast, which forecasts Earth embeddings at arbitrary lead times from irregularly sampled historical embeddings.
It can give both the Gaussian mean as a point forecast and ensemble forecasts (Figure 2), alongside the per-dimensional standard deviation to measure the uncertainty. To avoid assuming a fixed time step, we design a forecasting token space that turns each historical embedding into a token that carries its own acquisition date, and adds an adjustable query date to control the lead time. Using these tokens, we train a single transformer with two branches to produce Gaussian probability estimates and generated ensemble embeddings. Gaussian branch predicts a Gaussian mean, used as the point forecast, alongside the standard deviation directly estimated to measure the uncertainty. The embedding generation branch learns a rectified flow and draws realistic ensemble forecasts with different noise initializations. An example in Figure 1 shows the role of each branch and highlights the model’s flexible forecasting capability. With the forecasted embeddings, users can either compare them against observed embeddings to detect anomalies (shown later), or reuse the diverse pre-trained task heads to perform downstream analysis directly.
Figure 2. Overview of EForecast. (a) Each image in an irregular time series is converted into an Earth embedding by an Earth foundation model , which may differ across observations. (b) The historical embeddings with their dates, the location and the query date form a token sequence for a masked transformer . In one pass, the probability-estimation branch predicts the per-dimension Gaussian of the future embedding, and the embedding-generation branch denoises a noisy target by flow matching. (c) For forecasting, the probability branch gives a point forecast with uncertainty without sampling, and the generation branch turns noise draws into ensemble forecasts with Euler steps. (d) For anomaly change detection, a new observed embedding is compared with both the Gaussian forecast and the closest ensemble forecast in units of , giving an anomaly change map without change labels or a detection head.
By working in embedding space, EForecast is not tied to any one sensor or foundation model. We have tested 9 different embedding spaces such as OlmoEarth [24], Prithvi-EO-2.0 [61], AlphaEarth [7], Tessera [17] and the SatCLIP vision encoder [32], using both the Sentinel-1 and Sentinel-2 time series. With our approach, we observe consistent and large improvements in both forecast accuracy and uncertainty estimation. Remarkably, this improvement advances the SOTA in anomaly change detection without any label or head, simply by scoring a new observation by its distance to the forecast, thus raising the anomaly change detection from 41.58 to 48.70 on a global benchmark covering 19,801.60 km.
Our contributions are to:
forecast in Earth embedding space to alleviate low-level noise and focus on important information, which is a new paradigm enabled by geospatial foundation models.
design a flexible forecasting model EForecast that accepts diverse sensor embeddings, irregular historical observations, and arbitrary future date queries.
jointly train probability estimation and flow generation in a single transformer, giving an uncertainty-aware point forecast and ensemble forecasts.
perform accurate anomaly change detection directly from the forecasts without retraining the model.
Related Work
Earth analysis in embedding space
Earth embeddings compress a time series of repeated Earth observations into a high-level representation space [59]. They are produced by geospatial foundation models (GFMs), which are designed for different sensors, resolutions, and time-series settings. For example, CROMA [19] jointly encodes Sentinel-1 radar and Sentinel-2 optical imagery, DOFA [71] and HyperFree [34] processes images with varied channels, Scale-MAE [52] handles varying ground sampling distances, and Prithvi-EO-2.0 [61], Galileo [63] and OlmoEarth [24] ingest multi-temporal, multi-sensor inputs. More recently, some embeddings have been made available as precomputed global ready-to-use products [16], including Google’s AlphaEarth Foundations [7] and TESSERA [17] at 10 m, ESD [10] at 30 m, and Major TOM [14] at kilometre scale.
Both GFMs and embedding products operationalize the “embeddings as data” paradigm by reducing the burden of preprocessing heterogeneous observations and lowering computational and training-data requirements for downstream tasks. This paradigm has been widely adopted for diverse mapping tasks [11, 46, 54]. However, current studies primarily exploit the spatial continuity of embeddings; their temporal continuity, which is the focus of this work, remains largely unexplored[59].
Earth observation forecasting
Forecasting Earth observations has predominantly been for physical variables [4, 58, 68] and satellite imagery. Requena-Mesa et al. [53] introduced the task in 2021 with the classical EarthNet2021 benchmark, where future Sentinel-2 images are predicted using historical observations and future weather conditions. Earthformer [20] then applied cuboid space-time attention to this benchmark, and Contextformer [3] improved vegetation forecasting by fusing spatial context with weather drivers. EarthPT [57], instead, autoregressively forecasts pixel-level reflectance time series. More recently, generative models have been introduced to produce realistic future images, including diffusion models conditioned on past acquisitions [31], flow matching-based methods [74], and video generation-based models [23].
Despite their architectural and application differences, prior work on forecasting is typically in the original signal space, which is inevitably affected by observation-specific variabilities [59], such as clouds, shadows, and sensor noise [59]. We will show that although the forecasting images can look realistic, there are large errors when forecasts are converted into embedding space, implying loss of features.
Anomaly change detection
Not every change on the Earth’s surface is an anomaly. Time series forecasting can help identify normal changes such as crops growing, seasons turning, and human activity in unsupervised manner. For example, AnomalyCD [35] and ESIA [36] used long time-series observations to detect abrupt changes, where they assume that changes not seen before are anomalies. Some forecasting-based detection methods identify both gradual and abrupt changes [51, 64, 75]. However, their forecasts are still in the (original) spectral space and only apply to specific sensors.
Since forecasts provide a baseline for expected temporal variation, an accurate forecast can suppress uninteresting normal changes, so that the deviation of new observations from this baseline serves as an anomaly score. From this perspective, we hope that our improvements in forecasting will also benefit anomaly detection.
Earth embedding forecasting
This section discusses the details of our method, EForecast, which learns the pattern of changes in Earth embedding space, forecasts with a flexible time-series organization, and outputs uncertainty-aware ensemble outputs. Here we first formalise the embedding forecasting framework (§ 3.1), then instantiate it with a novel token space (§ 3.2). We carry out joint training of probability estimation and embedding generation (§ 3.3) and finally describe how we use inferences both for forecasting and for forecast-based anomaly change detection (§ 3.4).
A flexible ensemble framework for embedding forecasting
E²Forecast operates on embeddings already provided by existing Earth observation foundation models. It is agnostic to the foundation model type, missing values in the time-series, and target forecast date. Formally, let identify the foundation model that produces observation embedding . For a fixed imaging location , with longitude and latitude , a forecasting instance consists of a potentially heterogeneous embedding history and a future query date:
where is the observation context and is the observation date. Different embeddings may come from different foundation models depending on the demand. Multiple embeddings may also correspond to the same date. The query date specifies when to forecast, and we can also choose which target foundation model that the future embedding lies in.
The task is to model the conditional distribution of the future embedding :
In practice, many unobserved factors, such as human activity, can lead to different future states for the same observation context and query date. We therefore model it as an ensemble of multiple embeddings rather than a single regression output. Specifically, we introduce a Gaussian latent variable and a conditional generator , which maps each latent draw to a possible future embedding given and :
where resampling produces multiple future embeddings. However, sampling alone cannot answer the question of which ensemble member is the one we are most confident of, and how much uncertainty exists in each member of the ensemble. Existing work tackles this problem by postprocessing, say, 100 samples [15, 38], which is computationally inefficient. To increase efficiency, we pair with a parametric surrogate
whose mean is the point forecast and diagonal covariance provides the uncertainty information directly (§ 3.4). Together, the model reads a token sequence built from the observation context, the query date and a noised copy of the future embedding at flow time [37], and in a single forward outputs the Gaussian parameters and an estimate of the clean future embedding:
forward outputs the Gaussian parameters and an estimate of the clean future embedding:
where depend only on and are computed once per forecast, while estimates from and is re-evaluated at every solver step to generate samples from (§ 3.4).
Token space design
We convert all the elements of into a sequence of -dimensional tokens for batch transformer inputs. is decided by the choice of embedding space.
Location token. Seasonal changes and periodic activities vary across regions, and location information enables the model to learn these region-specific temporal patterns. We encode the imaging location with the random Fourier features (RFF) of GeoCLIP [9], chosen after comparing against other location encoders (§ 4). RFF has no trainable weights and is computed as follows:
where is the location in radians and is a frequency matrix with entries . The output in Eq. (6) is directly usable as a single token without any projection.
Date encoding and query date token. Observation dates provide essential temporal context for modelling how embeddings evolve. Here we represent dates using relative observation order for temporal order and day-of-year (DOY) for seasonality [17]. With the historical observations sorted by date, the -th one has and the forecast date has , which expresses its order relative to the forecast date, avoiding dependence on absolute calendar years. We use sine and cosine features to represent both the order across observations and variation within a year [24], allocating half of the dimensions to each. Each component contains sine–cosine pairs. Their scales are , , spaced geometrically from 1 to to encode both faster and slower temporal variation. The two components are
where is the seasonal phase and denotes concatenation. In Eq. (7), dividing by combines sensitivity to differences between adjacent observations with more slowly varying features for longer-term trends. In Eq. (8), multiplying the seasonal phase by instead provides frequencies from 1 to 100. The lowest frequency represents the annual cycle, while higher frequencies describe finer within-year variation. Concatenating Eqs. (7) and (8) gives . The query date is encoded in the same way and passed through its own learned projection to give the query date token .
Historical observation tokens. Although EForecast uses ready-to-use Earth embeddings, they need to be temporally ordered. Hence, we append a date encoding to each historical embedding before it is used as an input to the model. Two separate learned projections subsequently convert this input into a shared feature space, as described next. Let map embeddings from foundation model and let project the date encoding. Then, the resulting observation token is
Gaussian mean/std tokens. Since a generated future embedding is just one plausible future, we append two learned tokens to the transformer input: the Gaussian mean token and the Gaussian std token . Their transformer outputs are decoded by two linear heads into the parameters of the surrogate Eq. (4),
where are learnable linear heads and is a floor to keep the log-density of Eq. (11) bounded. These tokens, together with the location, history and query-date tokens, form the probability-estimation branch of the sequence (see § 2).
Embedding-generation tokens. Three further tokens serve the embedding-generation branch. A noisy target token carries a noised copy of the future embedding, and a flow time token carries its noise level . To prevent the model from stealing from historical observations when forecasting, we append a learned query token and use its output as the fore-
cast readout. It is initialised with small Gaussian noise, and its readout is mapped to the -dimensional target embedding space. These three tokens form the embedding-generation branch of the sequence (see 2).
Joint training of probability estimation and embedding generation
EForecast simultaneously trains both branches of the network. The probability-estimation branch directly predicts the mean and standard deviation of the future embedding distribution. The embedding-generation branch learns to draw plausible futures by flow matching. Both are read out from a single masked pass of Eq. (5).
Probability-estimation branch. The Gaussian tokens predict and via Eq. (10). We train them by minimising the negative log-likelihood of the future embedding under the surrogate Eq. (4):
The first term scales prediction errors by their estimated uncertainty, while prevents unbounded variance. This branch therefore provides a point forecast and uncertainty directly, without sampling or integration. Note that Gaussian NLL does not require the true conditional distribution to be Gaussian when learning the moments.
Embedding-generation branch. Rectified flow [40, 43] has a straight path from Gaussian noise to the target embedding . At a randomly sampled flow time , we construct
The model receives and through the noisy-target and flow-time tokens
where is a sinusoidal encoding. Since Earth embeddings are high-dimensional signals, the query-token output acts as a denoiser and predicts the clean future embedding rather than the velocity [37]; the path velocity then follows,
where is clipped at a small constant. We train the branch by matching the predicted velocity to the straight-path target :
Mixed objective. We jointly optimise the probability-estimation and embedding-generation branches as
Both branches are evaluated in one transformer pass, using an asymmetric mask that prevents probability-estimation tokens from attending to embedding-generation tokens. Both loss terms have the same scale, and equal weight.
Inference and anomaly change detection
The trained network supports forecasting and forecast-based anomaly change detection without retraining or task-specific heads.
Forecasting. Given any query date , EForecast accepts the available context and returns a point forecast, ensemble forecasts, and their predictive uncertainty as one tuple:
The Gaussian branch obtains in one pass with and . For each ensemble forecast, is advanced using Euler steps,
and all paths are processed in one batch. Thus, the same model can query arbitrary future dates, provide a fast point forecast with per-dimension uncertainty, and represent multiple possible futures when needed.
Anomaly change detection. After observing at , we compare it with both the Gaussian forecast and the closest ensemble forecast, scaled by the predicted uncertainty:
where division is element-wise. Taking the minimum, the score is high only when the observation is far from both the mean prediction and all ensemble forecasts, i.e. a change that any plausible forecast explains is treated as normal, while discounts expected variation. We use and at default; applying Eq. (19) across locations yields an anomaly map without change labels or a separately trained detection head.
Results
We evaluate EForecast on two tasks: forecasting future embeddings from irregular histories, and detecting anomalous changes by comparing new observations with these forecasts. We first describe the datasets, baselines, metrics, and implementation details (§ 4.1 to § 4.3). We then compare embedding-space and image-space forecasting and report anomaly change detection results (§ 4.4). Finally, § 4.5 reports related sensitivity and ablation results.
Datasets
S1_Embed for forecasting. This dataset forecasts embeddings of Sentinel-1 SAR time series under a cross-year condition. It contains 60 sites of km distributed over the globe (Figure 3a), 43 covering diverse biomes and 17 in large cities. Every descending Sentinel-1 RTC acquisition of these sites in 2024 and 2025 is converted into OlmoEarth-nano embeddings [24] with patch size 5, i.e. a grid of 128-dimensional vectors at 50 m. Models are trained on the 2024 series of all sites (1,991 scenes) and evaluated on 2025, with 10 sites (7 biome, 3 urban; 522 scenes) for validation and the remaining 50 sites (2,426 scenes) for testing. A total of 1M time-series samples, 20K per test site, are drawn to compute the metrics. Each sample forecasts with 7 earlier observations, which was chosen by the sensitivity analysis in Table 3. Notice that both the context intervals and the lead time to the target date are irregular (Figure 3a).

Figure 3. Site locations and irregular temporal sampling in the S1_Embed, S2_Embed and GSA test sets. Left: site locations. Middle: distribution of the lead time from the last context observation to the target date. Right: context dates (circles) and target date (star) of 25 random samples, relative to the target date. Context intervals and lead times vary across samples, spanning 1–170 days for S1_Embed, 5–60 days for S2_Embed and over a year for GSA, so models must forecast embeddings for an arbitrary target date.
| Historical number | 3 | 5 | 7 | 9 | 12 |
| Std | 1.198 | 1.162 | 1.138 | 1.131 | 1.134 |
| Ensemble Cosine | 0.1007 | 0.0945 | 0.0911 | 0.0902 | 0.0898 |
Table 3. Effect of the historical number on S1_Embed.
S2_Embed for forecasting. This dataset forecasts embeddings of optical multispectral time series under a cross-region condition. We use PASTIS [21], which contains 2,433 sites of pixels in four regions of France (Figure 3b), each observed by Sentinel-2 L2A on 38–61 dates spanning 2018–2019. Every date is encoded independently by Prithvi-EO-2.0-300M [61], yielding an grid of 1024-dimensional tokens that each cover pixels (160 m). Sites are split 6:2:2 into 1,461 training (69,066 scenes), 486 validation (23,273 scenes) and 486 test patches (22,883 scenes). A total of 100K time-series samples are drawn from the test patches to compute the metrics. A sample forecasts one token at a random target date from 37 randomly drawn earlier dates, again not necessarily consecutive, which yields irregular context intervals and lead times as well (Figure 3b).
We do not use the classical EarthNet2021 [53], as it provides only RGB and near-infrared bands, while existing foundation models need complete channels to generate dense embeddings.
GSA for anomaly change detection. GSA [36] is a global dataset for anomaly change detection. It covers 19,801.60 km and 55 events from six categories: dam break, burned area, fire point, marine debris, oil leak and water bloom. For each event, one post-event Sentinel-2 L2A image and a varying number of pre-event images form a monitoring stream, whose sites and irregular intervals are shown in Figure 3c, and the post-event image carries a pixel-level binary anomaly annotation. Anomalies are rare and small: most events have an anomaly ratio below 10%. We have tested EForecast with different embedding spaces on this dataset, and additionally collected annual Tessera 1.1 [17] and AlphaEarth [7] embeddings in the past three years as extra conditions.
Baselines and metrics
We compare EForecast with seven forecasting models that cover different output types: the deterministic Persistence, which takes the closest observation as the forecast, ETSformer [68] and Earthformer [20]; the probabilistic DeepAR [55] and Temporal Fusion Transformer (TFT) [39]; and the generative TMDM [38] and SimDiff [15]. Since none of these models was designed for embedding-space forecasting at an arbitrary date, we adapt them in two ways. In the image space, DeepAR, TFT and SimDiff forecast the image at the query date given our designed query-date token, and the forecast image is then converted into embeddings by the same foundation model for evaluation. In the embedding space, DeepAR, TFT, TMDM and SimDiff share the proposed token space with EForecast.
We evaluate forecasting performance from three aspects:
Uncertainty is assessed by the mean and standard deviation of the -score over all pixels and dimensions, whose ideal values are 0 and 1.
A point forecast is assessed by the point cosine distance between the predicted mean and the ground truth. For generative models, we take the mean of 100 samples for the point forecast.
Ensemble forecasts are assessed by the ensemble cosine distance, i.e. the smallest among the ensemble members, and by the CRPS [22]. When computing the ensemble cosine distance, we sample 100 samples for models that only output a Gaussian (DeepAR and TFT) and generate 10 forecasts for generative models (including EForecast). Flow and diffusion models use 10 Euler and 50 solver steps to generate one sample, respectively. Besides, we also compare with the persistence setting as a baseline, which takes the closest past observation as the forecast.
For anomaly change detection, using the GSA dataset, we compare with the 23 models benchmarked in [36]: 14 category-specific spectral-index detectors (e.g. dNBR and FAI), the bi-temporal change detectors CVA [5], SCG [60], MAD [49] and SAM-Bi [35], and the time-series detectors COLD [76], EOAD [8], PKAD [67], AnomalyCD [35] and ESIA [36]. Following [36], we report pixel-level precision, recall, and of the binary anomaly maps against the annotations.
Implementation details
The Transformer width equals the embedding dimension (128 for S1_Embed and GSA, 1024 for S2_Embed, 256 for the INR), so the location token enters without projection. Every input projection, including the separate projection of the query date, is a LayerNorm followed by a linear layer. The network has 8 pre-norm layers with 4 heads (8 for S2_Embed), an FFN width of four times the embedding dimension, GELU activations and dropout 0.05; the standard-deviation floor is . The flow branch operates on the target embedding scaled by the square root of its dimension, so each dimension has unit variance, and generated samples are scaled back and -normalised. Flow times are drawn from a logit-normal distribution with mean and standard deviation , and is clipped at 0.05. We train with AdamW at a constant learning rate of , no weight decay and gradient-norm clipping at 0.5. Forecasting models use 1000 steps per epoch: 600 epochs with batch size 256 on S1_Embed, 100 epochs with batch size 64 on S2_Embed, and 300 epochs with batch size 256 on GSA. The INR is trained for up to 300 epochs of samples with batch size 256. All results use .
Main results
Forecasting in the image space loses much of the semantic features. Since no prior model supports forecasting from irregular observations at an arbitrary lead time, we adapt existing baselines by equipping them with a query token for image forecasting, and with the proposed full token space for embedding forecasting. When evaluating image-space forecasting, we first use these models to forecast images and then convert the forecasts into embeddings. As shown in Table 1, most embedding-space models outperform their image-space counterparts. The gap is especially large in std and point forecast. The image-space baselines are strongly overconfident, with std between 1.73 and 4.07, whereas all embedding-space models stay closer to the ideal value of 1.0. The point forecast cosine distances are mostly over 0.17, much worse than the embedding ones, around 0.10.
| Method | S1_Embed Mean | S1_Embed Std | S1_Embed Point Cosine | S1_Embed Ensemble Cosine | S1_Embed CRPS | S2_Embed Mean | S2_Embed Std | S2_Embed Point Cosine | S2_Embed Ensemble Cosine | S2_Embed CRPS |
| Deterministic Models | ||||||||||
| Persistence | – | – | 0.1760 | 0.1760 | – | – | – | 0.1699 | 0.1699 | – |
| ETSformer [68] | – | – | 0.1109 | 0.1109 | – | – | – | 0.0737 | 0.0737 | – |
| Earthformer [20] | – | – | 0.1103 | 0.1103 | – | – | – | 0.0808 | 0.0808 | – |
| Image-Space Forecast with Designed Query-Date | ||||||||||
| DeepAR [55] | −0.0022 | 3.2313 | 0.1706 | 0.0926 | 0.0281 | −0.0184 | 1.7303 | 0.1812 | 0.2287 | 0.0073 |
| TFT [39] | 0.0018 | 4.0660 | 0.1698 | 0.1039 | 0.0311 | −0.0189 | 2.3262 | 0.2359 | 0.2539 | 0.0086 |
| SimDiff [15] | −0.0164 | 2.3214 | 0.2061 | 0.1491 | 0.0359 | −0.0012 | 2.0941 | 0.2454 | 0.1757 | 0.0072 |
| Embedding-Space Forecast with Designed Token Space | ||||||||||
| DeepAR* [55] | −0.0001 | 1.0771 | 0.1089 | 0.1439 | 0.0220 | 0.0005 | 1.0341 | 0.0679 | 0.0755 | 0.0045 |
| TFT* [39] | −0.0449 | 1.1776 | 0.1101 | 0.1367 | 0.0217 | −0.0424 | 1.0924 | 0.0951 | 0.1097 | 0.0052 |
| TMDM* [38] | 0.0062 | 0.6874 | 0.1210 | 0.3211 | 0.0308 | 0.0072 | 0.4925 | 0.0948 | 0.3359 | 0.0082 |
| SimDiff* [15] | −0.0046 | 1.4352 | 0.1105 | 0.1019 | 0.0236 | 0.0020 | 1.4991 | 0.0947 | 0.0876 | 0.0051 |
| E²Forecast (Ours) | −0.0005 | 1.1376 | 0.1081 | 0.0911 | 0.0213 | 0.0006 | 1.0250 | 0.0667 | 0.0613 | 0.0038 |
Table 1. Forecasting results on S1_Embed and S2_Embed. Metrics are defined in § 4.1. Deterministic models give a single forecast, so only cosine distances apply. * marks baselines substantially adapted to the embedding space. Bold: best; _underlined_: second best.
To illustrate this difference intuitively, we visualize two samples in Figure 4 and Figure 5. The forecasted images appear close to the targets in texture and color, but once converted into the embedding space, much of the high-level information is lost or altered. In S1_Embed, the converted embeddings collapse to a few dominant colors. In S2_Embed,

Figure 4. Image-space vs. embedding-space forecasting on S1_Embed. For each example, we show the target image with R=VV, G=VH, and B=VV-VH. Embeddings are visualized with the first three PCA components.

Figure 5. Image-space vs. embedding-space forecasting on S2_Embed. We show the RGB for the S2 image and visualize embeddings with the first three PCA components. The error maps are computed with error maps of over all embedding dimensions.
their PCA maps even resemble the target, but the error maps on all dimensions show large differences in the other principal components, while the error of EForecast remains close to zero.
Learning normal change in the embedding space generalizes across anomaly categories. As shown in Table 2, EForecast achieves the best average precision and , raising from 41.58 of ESIA to 48.70, and it gives the best in five of the six categories. The time-series detectors COLD and PKAD obtain higher average recall, but their precision stays below 30, i.e. they flag much of the normal variation as anomalies. Spectral indices are strong only on the category they were designed for, such as dNDVI on burned area. EForecast instead learns what normal change looks like in the embedding space, so the same model works across all categories without category-specific design. Marine debris is the exception, where EForecast ranks third in : the debris covers only a few pixels, so even small false alarms, such as the scattered high scores along the coastline in Figure 6, noticeably lower its precision. In the other categories, the forecast embedding resembles the observed one everywhere except in the event region, so the anomaly score concentrates on the annotated anomalies.
| Model | Dam break R | Dam break P | Dam break F | Model | Burned area R | Burned area P | Burned area F | Model | Fire point R | Fire point P | Fire point F | Input type | ||
| Specialized anomaly localization models | ||||||||||||||
| MNDWI [72] | 74.64 | 37.33 | 40.37 | dNDVI [29] | 39.01 | 87.32 | 48.24 | TAI [70] | 2.91 | 14.81 | 3.92 | Reflectance | ||
| MBWI [65] | 62.59 | 42.52 | 41.64 | dNBR [45] | 43.30 | <u>82.05</u> | 52.50 | NDFI [42] | 24.64 | <u>49.69</u> | 21.72 | Reflectance | ||
| AWEI [28] | 69.21 | 43.73 | 43.44 | RdNBR [48] | 63.97 | 67.10 | 62.13 | AFD-S2 [27] | 2.33 | 89.95 | 4.51 | Reflectance | ||
| General change detection models (bi-temporal) | ||||||||||||||
| CVA [5] | 33.65 | 58.90 | 29.82 | CVA [5] | 15.90 | 71.42 | 23.56 | CVA [5] | 40.68 | 36.97 | 31.35 | Reflectance | ||
| SCG [60] | 62.83 | 43.04 | 38.97 | SCG [60] | 38.36 | 62.01 | 42.11 | SCG [60] | 60.21 | 33.05 | 31.87 | Reflectance | ||
| MAD [49] | 48.48 | 56.05 | 35.77 | MAD [49] | 26.78 | 71.66 | 35.73 | MAD [49] | 52.16 | 34.30 | <u>33.22</u> | Reflectance | ||
| SAM-Bi [35] | 65.94 | 46.43 | 43.76 | SAM-Bi [35] | 50.04 | 56.23 | 47.18 | SAM-Bi [35] | 48.13 | 22.58 | 22.62 | SAM features | ||
| General anomaly localization models (time-series) | ||||||||||||||
| COLD [76] | 91.06 | 39.27 | 44.71 | COLD [76] | 58.45 | 64.39 | 47.01 | COLD [76] | 67.62 | 23.69 | 22.79 | Reflectance | ||
| EOAD [8] | 52.21 | 38.15 | 36.44 | EOAD [8] | 19.98 | 52.21 | 24.54 | EOAD [8] | 53.33 | 29.56 | 30.41 | Vegetation index | ||
| PKAD [67] | 71.15 | 39.58 | 28.69 | PKAD [67] | 89.31 | 57.81 | 66.97 | PKAD [67] | 62.68 | 17.35 | 20.27 | Vegetation index | ||
| AnomalyCD [35] | 56.70 | 66.20 | 52.24 | AnomalyCD [35] | 46.92 | 68.16 | 49.38 | AnomalyCD [35] | 38.37 | 35.79 | 29.63 | SAM features | ||
| ESIA [36] | 62.98 | 58.77 | <u>54.32</u> | ESIA [36] | <u>82.36</u> | 56.04 | 63.31 | ESIA [36] | 41.60 | 31.06 | 32.09 | HyperFree features | ||
| EForecast | <u>87.27</u> | <u>60.46</u> | 60.91 | EForecast | 80.76 | 81.62 | 78.38 | EForecast | 48.03 | 41.24 | 38.70 | Embeddings | ||
| Model | Marine debris R | Marine debris P | Marine debris F | Model | Oil leak R | Oil leak P | Oil leak F | Model | Water bloom R | Water bloom P | Water bloom F | Average R | Average P | Average F |
| Specialized anomaly localization models | ||||||||||||||
| FDI [2] | 62.08 | 20.93 | 19.44 | NDOI [6] | 51.29 | 4.34 | 6.16 | FAI [26] | 57.36 | 28.76 | 34.30 | – | – | – |
| PI [62] | 22.68 | 5.89 | 6.38 | – | – | – | – | SABI [1] | 61.21 | 39.62 | 36.93 | – | – | – |
| General change detection models (bi-temporal) | ||||||||||||||
| CVA [5] | 46.35 | 7.81 | 6.56 | CVA [5] | 8.61 | 4.65 | 6.03 | CVA [5] | 12.15 | 45.09 | 19.09 | 26.22 | 37.47 | 19.40 |
| SCG [60] | 84.77 | 0.73 | 1.43 | SCG [60] | 43.05 | 6.45 | 10.52 | SCG [60] | 52.15 | 29.72 | 29.91 | 56.90 | 29.17 | 25.80 |
| MAD [49] | 52.43 | 4.82 | 5.13 | MAD [49] | 10.16 | 4.41 | 6.10 | MAD [49] | 19.23 | 42.07 | 26.30 | 34.87 | 35.55 | 23.71 |
| SAM-Bi [35] | 18.09 | 10.42 | 11.60 | SAM-Bi [35] | 53.37 | 11.45 | 18.47 | SAM-Bi [35] | 41.25 | 18.20 | 23.10 | 46.14 | 27.55 | 27.79 |
| General anomaly localization models (time-series) | ||||||||||||||
| COLD [76] | <u>70.34</u> | 4.55 | 4.48 | COLD [76] | 69.64 | 3.90 | 7.21 | COLD [76] | 51.79 | 33.61 | 33.91 | <u>68.15</u> | 28.24 | 26.69 |
| EOAD [8] | 42.85 | 0.40 | 0.78 | EOAD [8] | 32.62 | 10.38 | 12.91 | EOAD [8] | 38.90 | 31.49 | 28.14 | 39.98 | 27.03 | 22.20 |
| PKAD [67] | 67.29 | 0.99 | 1.83 | PKAD [67] | <u>66.18</u> | 4.23 | 7.80 | PKAD [67] | 83.95 | 19.80 | 28.57 | 73.43 | 23.29 | 25.69 |
| AnomalyCD [35] | 28.07 | 10.09 | 14.36 | AnomalyCD [35] | 44.01 | 7.63 | 12.87 | AnomalyCD [35] | 55.50 | 16.49 | 22.72 | 44.93 | 34.06 | 30.27 |
| ESIA [36] | 34.56 | 21.26 | <u>18.33</u> | ESIA [36] | 38.73 | <u>18.11</u> | <u>20.03</u> | ESIA [36] | 69.80 | <u>55.30</u> | 61.53 | 54.99 | <u>40.05</u> | <u>41.58</u> |
| EForecast | 56.55 | 12.04 | 14.61 | EForecast | 42.87 | 23.27 | 26.01 | EForecast | <u>77.38</u> | 74.39 | 73.58 | 65.48 | 48.84 | 48.70 |
Table 2. Anomaly change detection results on GSA. Pixel-level recall (R), precision (P) and F (%) for each of the six event categories and their average. Baseline results are taken from [36]. Best and second best per column are in bold and <u>underlined</u>.

Figure 6. Qualitative anomaly change detection results on GSA. One event per category. (a) Post-event Sentinel-2 RGB image with the annotated anomaly outlined in red. (b) Embedding forecast by EForecast and (c) embedding observed on the post-event date, both projected to RGB. (d) Anomaly score of Eq. (19).
Discussion
We now place our results in a broader context and quantify the benefits from our design choices.
What benefit does an ensemble forecast bring? Unlike deterministic models, EForecast generates ensemble forecasts to account for factors that the context cannot predict. However, more forecasts do not necessarily mean a better forecast, as the probabilistic DeepAR and TFT in Table 1 show. In a low-dimensional space, such as the two VV and VH bands of an S1 image, drawing 100 samples from the predicted Gaussian does improve DeepAR’s point forecast—from 0.1706 to 0.0926—and TFT’s point forecast—from 0.1698 to 0.1039—in cosine distance. However, as the dimensionality increases, such as to the ten bands of an S2 image, the same sampling worsens the forecast, from 0.1812 to 0.2287 (DeepAR) and from 0.2359 to 0.2539 (TFT). Indeed, in the 128- and 1024-dimensional embedding spaces, the best of 100 Gaussian samples is worse than the point forecast of the same model on both datasets. We hypothesize that the reason for this effect is that, assuming dimensions are independent, a sample drifts further from the mean as the dimension grows, so both DeepAR and TFT struggled to generate realistic embeddings via direct sampling.
Generative models avoid this issue by producing whole embeddings without a distribution assumption, but, nevertheless, not all of them succeed. For example, TMDM* appears to predicts the noise with its ensemble cosine distance rising from 0.1210 to 0.3211 and from 0.0948 to 0.3359 (denoising models of this type are known to struggle with high-dimensional signals [37]). Instead, SimDiff* and EForecast follow the -prediction style that outputs the embedding directly rather than noises [37], and both improve performance from an ensemble with only 10 members: SimDiff* from 0.1105 to 0.1019 on S1_Embed, and EForecast from 0.1087 to 0.0911.
To show this benefit more clearly, we visualize the forecasts of EForecast for random test samples of S1_Embed in Figure 7 (the corresponding S2_Embed in Figure 11 of Sec. A.1). The point forecast follows the overall trend of the target but misses many individual dimensions, whereas the closest ensemble member reproduces almost every dimension, reducing from 0.32–0.44 to 0.04–0.07 on S1_Embed and from 0.58–0.65 to 0.16 on S2_Embed. Projected into 2D, the five ensemble members spread around the point forecast, which sits near their centre, and form a local manifold of plausible futures.

Figure 7. Point vs. ensemble forecasts of EForecast on S1_Embed. Each column is a random test sample. Top: the point forecast against the ground truth over the 128 dimensions. Middle: the ensemble member closest to the ground truth. Bottom: five ensemble members projected onto the first two PCA components of the test targets of the same site. is the cosine distance to the ground truth.

Figure 11. Point vs. ensemble forecasts of E²Forecast on S2_Embed. Same layout as Figure 7, over the 1024 dimensions. is the cosine distance to the ground truth.
What if longer historical observations are provided?
We train and test EForecast on S1_Embed with 3, 5, 7, 9 and 12 historical observations per series. As before, every series has its own irregular intervals and lead time. As Table 3 shows, a longer history gives a more accurate estimate of the uncertainty: std falls stably. It is therefore not surprising that S2_Embed, with 37 historical observations, reaches a std of 1.0250 (Table 1). Besides, more prior information does not suppress the generative ability, rather, it makes the ensemble more realistic, and the ensemble cosine distance decreases from 0.1007 to 0.0898. We find that beyond 7 observations both metrics change only marginally, so we use 7 historical observations for S1_Embed.
What is the effect of date and location tokens?
Recall that S1_Embed covers sites across the globe. To study the effect of location and date, we ablate the corresponding tokens. The date token tested here is the date encoding added to each historical observation token in Eq. (9). Note that the query date token is necessary for the forecast and is therefore always kept. Besides removing the location token, we also replace RFF with other location encoders: the spherical
harmonics (SH) of SatCLIP [32] and the spherical harmonics with discretized-DFS (SHDD) of LocDiffusion [69]. As Table 4 shows, removing either token hurts the performance. Removing the date token matters more: it is the only setting with a std above 1.2 (1.2324) and an ensemble cosine distance above 0.10 (0.1008). Among the location encoders, SH and RFF perform better than SHDD. Considering the large difference in std , 1.1382 against 1.1627, we choose RFF as our location encoder.
Table 4. Ablation of the date and location tokens on S1_Embed. “w/o date token” removes the date encoding added to the historical observation tokens (Eq. (9)); The bottom block compares location encoders.
What is the effect of the forecast lead time?
We group the test samples of both forecast datasets by the lead time from the last historical observation to the target date and evaluate each group separately (Figure 12 in Sec. A.2). On both datasets, the uncertainty estimate remains stable as the lead time grows: mean stays within , and std stays between 1.12 and 1.21 on S1_Embed and close to 1 on S2_Embed (Figure 12a,Figure 12d). The forecast itself, however, gradually degrades with lead times. Both the point and the ensemble cosine distance rise (Figure 12b,Figure 12e), as does the CRPS (Figure 12c,Figure 12f); on S2_Embed the CRPS almost doubles, from 0.0033 at 5 days to 0.0060 at 37 days. One exception is the 11-day group of S1_Embed, which is even better than the 4-day group, with a point cosine distance of 0.1045 against 0.1125. We attribute this to its size: with 532k items, it holds several times more samples than other groups, so the model is most familiar with this lead time.

Figure 12. Effect of the forecast lead time. Test samples of S1_Embed (a–c) and S2_Embed (d–f) are grouped by the lead time, and the tick labels give the mean lead time and the number of items of each group. (a, d) Mean and std of the -score of the Gaussian head; (b, e) cosine distance of the point forecast and of the best of ensemble members; (c, f) CRPS of the Gaussian head. Error bars are 95% confidence intervals of the group mean, and dashed lines mark the value over the whole test set.
Is E2Forecast compatible with different embedding spaces? We have already shown that E2Forecast works in the OlmoEarth and Prithvi embedding spaces on S1_Embed and S2_Embed. We additionally test seven embedding spaces on GSA (Figure 8; full results in Table 6 of Sec. A.4): OlmoEarth [24], Galileo [63], CROMA [19], SatMAE [13] and Clay [12], and OlmoEarth combined with the annual Tessera or AlphaEarth embeddings. Since anomalies are annotated per pixel, we only use foundation models that support a patch size no larger than 8, which leaves out those with a patch size of 16, such as Prithvi-EO-2.0 [61] and DOFA [71]. With the same model and training recipe, six of the seven spaces exceed the best baseline ESIA [36] (41.58 F1), showing that E2Forecast does not depend on a particular embedding space. In general, a stronger foundation model brings better results. For example, the recent multi-temporal, multi-sensor OlmoEarth and Galileo reach 48.70 and 46.53 F1, ahead of the older Clay, SatMAE and CROMA models (43.64, 42.61 and 36.82). Besides, different embedding spaces can also be used together. We find that adding Tessera v1.1 to OlmoEarth gives the best result, raising precision to 49.65 and F1 to 49.29.
Figure 8. Average F1 of E2Forecast on GSA with different embedding spaces. Hatched bars add annual Tessera or AlphaEarth embeddings as extra conditions to OlmoEarth, and the dashed line marks ESIA, the best baseline in Table 2. Recall and precision are given in Table 6.
Table 6. Results of Forecast on GSA with different embedding spaces. Average pixel-level recall (R), precision (P) and (%) over the six event categories. “+” adds annual embeddings to OlmoEarth as extra conditions. Best and second best per column are in bold and underlined.
Connection with implicit neural representations
In this section, we explore the connection between implicit neural representations and E2Forecast. An implicit neural representation (INR) stores a signal as a network that maps a coordinate to the value there [56]. With an empty context ( in Eq. (1)), E2Forecast maps the location token alone to an embedding at , i.e. it becomes an INR of the global field of foundation-model embeddings.
To test the performance of such an INR, we compare E2Forecast with an empty context with SatCLIP [32]. To do a fair comparison, we train on the same Sentinel-2 embeddings as SatCLIP. At the inference stage, we used our trained INR model to generate semantic embeddings for each unseen location, and then tuned an MLP layer to regress it onto the five tasks of population density, elevation, air temperature [25, 32] and California median house value and income [50], reporting the test . For the two California Housing tasks, five z-scored tabular features (House-Age, AveRooms, AveBedrms, Population, AveOccup) are appended to the embedding for all methods. Since SatCLIP does not release its downstream splits, and a random split may leak information, we therefore split each dataset into disjoint train/val/test regions (30/10/60 bands along latitude or longitude, with a guard gap, as shown as the top row of Figure 13 in Sec. A.3).

Figure 13. Held-out predictions on all five tasks. Top row: spatial split (train, val, test). Below: ground truth and each method on the test band only, sharing the ground-truth colour scale per column; both SatCLIP rows use the ViT-16 tower. Axes are longitude and latitude in degrees.
We compare our approach with with SatCLIP (all four released checkpoints, ResNet-18 or ViT-16 with degree or 40), as well as GeoCLIP, CSP [47], GPS2Vec [73] and raw coordinates (Table 5). Under region-disjoint splits, every deterministic encoder exhibited degraded performance, with eight of nine baselines have a negative average . The best, GeoCLIP, reaches an average of only 0.0593, because the probe extrapolates a fixed spherical basis rather than the target. E2Forecast is the only method positive on all five tasks and raises the average to 0.4569. GeoCLIP leads only on the two California tasks, where the geographic shift between the train and test spans fewer degrees than the other tasks. In Figure 13 (Sec. A.3), E2Forecast reproduces the large-scale spatial patterns of population, relief, and temperature in the held-out regions.
| Task | Identity | GPS2Vec | CSP-iNat | CSP-FMoW | RN18 | RN18 | ViT-16 | ViT-16 | GeoCLIP | Forecast |
| Population | −2.9862 | −0.5996 | −2.0198 | −0.1886 | 0.1468 | 0.2373 | −0.1801 | 0.1028 | 0.3717 | 0.6064 |
| Elevation | −6.8305 | −0.1472 | −4.4808 | −0.5511 | −19.2185 | −3.6502 | −3.4695 | −0.0937 | −1.9232 | 0.1687 |
| Temperature | −94.6671 | −0.2646 | −0.6582 | 0.1980 | 0.2065 | 0.5840 | 0.3687 | 0.5450 | 0.7094 | 0.7628 |
| House value | −0.0118 | −0.0316 | 0.0774 | −0.5702 | 0.2182 | −0.2985 | 0.0083 | −2.8809 | 0.4740 | 0.2183 |
| Median income | −0.0185 | 0.5964 | 0.1035 | −1.8341 | 0.1644 | 0.0224 | 0.5032 | 0.2731 | 0.6646 | 0.5282 |
| Average | −20.9028 | −0.0893 | −1.3956 | −0.5892 | −3.6965 | −0.6210 | −0.5539 | −0.4107 | 0.0593 | 0.4569 |
Table 5. Results when adapting Forecast as INR on five regression tasks (). All methods share the same probe and spatial splits. : Legendre degree of SatCLIP. Best per row in bold.
For a visual comparison, we show the generated global-scale embeddings in Figure 9. Each encoder is queried on a global grid, and embeddings are projected to RGB beside a 2025 land-cover map [30]. Although it never saw land-cover labels, E2Forecast produces a field that follows coastlines and ice sheets and whose regions match land cover: the Sahara and Arabian deserts, the Amazon and Congo forests, and the boreal belt are clearly visible. In contrast, SatCLIP, which is aligned to the same embeddings using a contrastive loss, shows only a smooth low-frequency function of the coordinate, which explains why its probe does not extrapolate well. Similarly, GeoCLIP adds speckle and polar streaks unrelated to the surface.

Figure 9. The global embedding field compared with different INR models. Embeddings queried on a grid, projected to RGB by PCA, with the ocean masked. Left: the 2025 Esri land-cover map [30].
Because the S2-100K dataset used in Figure 9 has no acquisition dates, we further train Forecast with the same SatCLIP vision encoder on the SSL4EO-S12 dataset [66], which samples four seasons per location, and enables the query date token to generate a field for each month (Figure 10). After removing the annual mean, the field shows a clear yearly cycle: December and June are near opposite states, and the Northern and Southern Hemispheres change in antiphase, e.g., boreal regions and Australia swap colors between winter and summer. This strongly suggests that Forecast has learned both regional and seasonal structure.

Figure 10. Time- and location-aware global embedding field. Forecast trained on SSL4EO-S12 [66] with the query date token, queried on a grid for each month. The annual mean is subtracted at each location before projecting to RGB.
Conclusion
We present Forecast that makes forecasts in the embedding space of Earth foundation models and find that it is more effective than prior approaches that forecast pixels in the native spectral space. Its key innovation is a token space that allows it to take as input irregular histories from different sensors and foundation models and forecast at any query date. Across Sentinel-1 and Sentinel-2 benchmarks, adapting existing forecasters into embedding space consistently improves their accuracy and calibration as in Table 1, with Forecast resulting in the best point and ensemble forecasts. Interestingly, the same forecasts also detect anomalous change without labels or retraining. Moreover, with an empty context, the model becomes a powerful implicit neural representation of the global embedding field.
We acknowledge that Forecast still has two limitations. First, the probability-estimation branch ignores the correlation between embedding dimensions. This is because the surrogate in Eq. (4) uses a diagonal covariance, predicting the mean and standard deviation of each dimension independently, whereas the dimensions of foundation-model embeddings are usually highly correlated. We tried extending to model this cross-dimensional correlation, but it substantially degraded performance. We believe this reflects our particular formulation rather than the idea itself, and leave a better way of modelling these correlations to future work. Second, Forecast forecasts only from past embeddings, location and date, without accounting for the external factors that drive land surface change, such as weather, which govern vegetation growth, floods and fires, and terrain, which shapes how water and land cover evolve. Adding such Earth system drivers as extra condition tokens is a natural next step that we intend to explore in future work.
Additional results
Point vs. ensemble forecasts on S2_Embed
Figure 11 repeats the visualisation of Figure 7 for random test samples of S2_Embed.
Effect of the forecast lead time
Figure 12 reports the metrics of Forecast per lead-time group on S1_Embed and S2_Embed, as discussed in § 4.5.
Held-out predictions of the location-query INR
Figure 13 shows the spatial splits and the held-out predictions of each location encoder on the five regression tasks of § 5.
Anomaly change detection with different embedding spaces
Table 6 gives the average recall, precision and of Forecast on GSA with each embedding space plotted in Figure 8, as discussed in § 4.5.
References
- [1]Fahad Alawadi. Detection of surface algal blooms using the newly developed algorithm surface algal bloom index (SABI). In Remote Sensing of the Ocean, Sea Ice, and Large Water Regions, pages 45–58. SPIE, 2024.
- [2]Mattia Ali, Francesco Razzano, Sergio Vitale, Giampaolo Ferraioli, Vito Pascazio, Gilda Schirinzi, and Silvia Ullo. Assessment of spectral-based solutions for the detection of floating marine debris. In International Conference on Networks, Multimedia and Information Technology (NMIT-CON), pages 1–5, 2024.arxiv.org/abs/2408.10187
- [3]Vitus Benson, Claire Robin, Christian Requena-Mesa, Lazaro Alonso, Nuno Carvalhais, José Cortés, Zhihan Gao, Nora Linscheid, Mélanie Weynants, and Markus Reichstein. Multi-modal learning for geospatial vegetation forecasting. In CVPR, pages 27788–27799, 2024.arxiv.org/abs/2303.16198
- [4]Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiao-tao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619:533–538, 2023.DOI
- [5]Francesca Bovolo and Lorenzo Bruzzone. A theoretical framework for unsupervised change detection based on change vector analysis in the polar domain. IEEE Transactions on Geoscience and Remote Sensing, 45(1):218–236, 2007.DOI
- [6]Camilla Brekke and Anne H. S. Solberg. Oil spill detection by satellite remote sensing. Remote Sensing of Environment, 95:1–13, 2005.DOI
- [7]Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha Samsikova, Chenhui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko, Noel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan, Rebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. AlphaEarth Foundations: An embedding field model for accurate global mapping from sparse label data. arXiv preprint arXiv:2507.22291, 2025.arxiv.org/abs/2507.22291
- [8]Liliana Castillo-Villamor, Andy Hardy, Pete Bunting, Wilson Llanos-Peralta, Miguel Zamora, Yeison Rodriguez, and Diego Alejandro Gomez-Latorre. The Earth Observation-based Anomaly Detection (EOAD) system: A simple, scalable approach to mapping in-field and farm-scale anomalies using widely available satellite imagery. International Journal of Applied Earth Observation and Geoinformation, 104, 2021.DOI
- [9]Vicente Vivanco Cepeda, Gaurav Kumar Nayak, and Mubarak Shah. GeoCLIP: Clip-inspired alignment between locations and images for effective worldwide geo-localization. In NeurIPS, 2023.arxiv.org/abs/2309.16020
- [10]Shuang Chen, Jie Wang, Shuai Yuan, Jiayang Li, Yu Xia, Yuanhong Liao, Junbo Wei, Jincheng Yuan, Xiaoqing Xu, Xiaolin Zhu, Peng Zhu, Hongsheng Zhang, Yuyu Zhou, Haohuan Fu, Huabing Huang, Bin Chen, Fan Dai, and Peng Gong. Democratizing planetary-scale analysis: an ultralightweight Earth embedding database for accurate and flexible global land monitoring. Earth System Science Data, 18:5375–5398, 2026.arxiv.org/abs/2601.11183
- [11]Yusen Cheng, Qin Feng Zhu, and Lei Fan. From landslide conditioning factors to satellite embeddings: Evaluating the utilisation of Google AlphaEarth for landslide susceptibility mapping using deep learning. arXiv preprint arXiv:2601.07268, 2026.arxiv.org/abs/2601.07268
- [12]Clay Foundation. Clay: A foundation model for Earth observation. https://github.com/Clay-foundation/model, 2024.DOI
- [13]Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In NeurIPS, 2022.arxiv.org/abs/2207.08051
- [14]Mikolaj Czerkawski, Marcin Kluczek, and Jedrzej S. Bojanowski. Global and dense embeddings of Earth: Major TOM floating in the latent space. arXiv preprint arXiv:2412.05600, 2024.arxiv.org/abs/2412.05600
- [15]Hang Ding, Xue Wang, Tian Zhou, and Tao Yao. SimDiff: Simpler yet better diffusion model for time series point forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 40(25):20781–20789, 2026.arxiv.org/abs/2511.19256
- [16]Heng Fang, Adam J. Stewart, Isaac Corley, Xiao Xiang Zhu, and Hossein Azizpour. Earth embeddings as products: Tax- onomy, ecosystem, and standardized access. arXiv preprint arXiv:2601.13134, 2026.
- [17]Zhengpeng Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young, Madeline C. Lisa ius, Markus Immitzer, Toby Jackson, James Ball, David A. Coomes, Anil Madhavapeddy, Andrew Blake, and Srinivasan Keshav. TESSERA: Temporal embeddings of surface spectra for Earth representation and analysis. In CVPR, 2026.arxiv.org/abs/2506.20380
- [18]Zhengpeng Feng, Sadiq Jaffer, Ira Shokar, Jovana Knezevic, James Ball, Pedro Sousa, Mark Elvers, Madeline Lisa ius, Clement Atzberger, Robin Young, et al. Tessera v2: Scaling pixel-wise earth foundation models. arXiv preprint arXiv:2607.03949, 2026.arxiv.org/abs/2607.03949
- [19]Anthony Fuller, Koreen Millard, and James R. Green. CROMA: Remote sensing representations with contrastive radar-optical masked autoencoders. In NeurIPS, 2023.arxiv.org/abs/2311.00566
- [20]Zhihan Gao, Xingjian Shi, Hao Wang, Yi Zhu, Yuyang Wang, Mu Li, and Dit-Yan Yeung. Earthformer: Exploring space-time transformers for Earth system forecasting. In NeurIPS, 2022.arxiv.org/abs/2207.05833
- [21]Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, pages 4872–4881, 2021.arxiv.org/abs/2107.07933
- [22]Tilmann Gneiting and Adrian E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477):359–378, 2007.
- [23]Yuchao Gu, Weijia Mao, and Mike Zheng Shou. Long-context autoregressive video modeling with next-frame prediction. arXiv preprint arXiv:2503.19325, 2025.arxiv.org/abs/2503.19325
- [24]Henry Herzog, Favyen Bastani, Yawen Zhang, Gabriel Tseng, Joseph Redmon, Hadrien Sablon, Ryan Park, Jacob Morrison, et al. OlmoEarth: Stable latent image modeling for multimodal Earth observation. arXiv preprint arXiv:2511.13655, 2025.arxiv.org/abs/2511.13655
- [25]Josh Hooker, Gregory Duvellier, and Alessandro Cescatti. A global dataset of air temperature derived from satellite remote sensing and weather stations. Scientific Data, 5:180246, 2018.DOI
- [26]Chuanmin Hu. A novel ocean color index to detect floating algae in the global oceans. Remote Sensing of Environment, 113:2118–2129, 2009.DOI
- [27]Xikun Hu, Yifang Ban, and Andrea Nascetti. Sentinel-2 MSI data for active fire detection in major fire-prone biomes: A multi-criteria approach. International Journal of Applied Earth Observation and Geoinformation, 101:102347, 2021.DOI
- [28]Hao Jiang, Min Feng, Yunqiang Zhu, Ning Lu, Jianxi Huang, and Tong Xiao. An automated method for extracting rivers and lakes from Landsat imagery. Remote Sensing, 6:5067–5089, 2014.
- [29]Zhangyan Jiang, Alfredo R. Huete, Jin Chen, Yunhao Chen, Jing Li, Guangjian Yan, and Xiaoyu Zhang. Analysis of NDVI and scaled difference vegetation index retrievals of vegetation fraction. Remote Sensing of Environment, 101:366–378, 2006.DOI
- [30]Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C. Mazzariello, Mark Mathis, and Steven P. Brumby. Global land use/land cover with Sentinel 2 and deep learning. In IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 4704–4707, 2021.DOI
- [31]Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B. Lobell, and Stefano Ermon. DiffusionSat: A generative foundation model for satellite imagery. In ICLR, 2024.arxiv.org/abs/2312.03606
- [32]Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. SatCLIP: Global, general-purpose location embeddings with satellite imagery. In AAAI, 2025.arxiv.org/abs/2311.17179
- [33]Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Scheider, Jasper Zschiegner, Hao Wang, and Yuyang Wang. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. In Advances in Neural Information Processing Systems, pages 28341–28364, 2023.arxiv.org/abs/2307.11494
- [34]Jingtao Li, Yingyi Liu, Xinyu Wang, Yunning Peng, Chen Sun, Shaoyu Wang, Zhendong Sun, Tian Ke, Xiao Jiang, Tangwei Lu, Anran Zhao, and Yanfei Zhong. HyperFree: A channel-adaptive and tuning-free foundation model for hyperspectral remote sensing imagery. In CVPR, 2025.arxiv.org/abs/2503.21841
- [35]Jingtao Li, Yanfei Zhong, Qian Zhu, Xinyu Wang, and Liangpei Zhang. Unsupervised anomaly localization with high-resolution and time-series satellite imagery: From global disaster dataset to zero-shot model via SAM. ISPRS Journal of Photogrammetry and Remote Sensing, 237:467–491, 2026.DOI
- [36]Jingtao Li, Qian Zhu, Xinyu Wang, Deren Li, Liangpei Zhang, and Yanfei Zhong. Earth surface immune system for rapid monitoring of unknown anomalies. arXiv preprint arXiv:2609.20662, 2026.arxiv.org/abs/2609.20662
- [37]Tiancheng Li and Kaiming He. Back to basics: Let denoising generative models denoise. arXiv preprint arXiv:2511.13720, 2025.arxiv.org/abs/2511.13720
- [38]Yuxin Li, Wenchao Chen, Xinyue Hu, Bo Chen, Baolin Sun, and Mingyuan Zhou. Transformer-modulated diffusion models for probabilistic multivariate time series forecasting. In International Conference on Learning Representations, 2024.
- [39]Bryan Lim, Sercan Ö. Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4):1748–1764, 2021.arxiv.org/abs/1912.09363
- [40]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ICLR, 2023.arxiv.org/abs/2210.02747
- [41]Hao Liu, Xiaoxing Zhang, Chun Yang, and Xiaobin Zhu. Semantic-enhanced time-series forecasting via large language models. In International Conference on Learning Representations, 2026.arxiv.org/abs/2508.07697
- [42]Jinglong Liu, Feng Zhao, Yunjia Wang, Yanan Wang, Sen Du, Lin Dang, and Jordi J. Mallorqui. Advancing coal fire detection model for large-scale areas based on RS indices and machine learning. International Journal of Applied Earth Observation and Geoinformation, 140:104587, 2025.DOI
- [43]Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023.arxiv.org/abs/2209.03003
- [44]Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Timer-XL: Long-context transformers for unified time series forecasting. In International Conference on Learning Representations, 2025.arxiv.org/abs/2410.04803
- [45]Duncan C. Lutes, Robert E. Keane, John F. Caratti, Carl H. Key, Nathan C. Benson, Steve Sutherland, and Larry J. Gangi. FIREMON: Fire effects monitoring and inventory system. Technical Report RMRS-GTR-164, U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, 2006.DOI
- [46]Yuchi Ma, Yawen Shen, Anu Swatantran, and David B. Lobell. Harvesting AlphaEarth: Benchmarking the geospatial foundation model for agricultural downstream tasks. arXiv preprint arXiv:2601.00857, 2026.DOI
- [47]Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. CSP: Self-supervised contrastive spatial pre-training for geospatial-visual representations. In Int. Conf. Mach. Learn., 2023.arxiv.org/abs/2305.01118
- [48]Jay D. Miller and Andrea E. Thode. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta normalized burn ratio (dNBR). Remote Sensing of Environment, 109:66–80, 2007.DOI
- [49]Allan Aasbjerg Nielsen. The regularized iteratively reweighted MAD method for change detection in multi- and hyperspectral data. IEEE Transactions on Image Processing, 16(2):463–478, 2007.pubmed.ncbi.nlm.nih.gov/17269639
- [50]R. Kelley Pace and Ronald Barry. Sparse spatial autoregressions. Statistics & Probability Letters, 33(3):291–297, 1997.
- [51]Amy H. Pickens, Matthew C. Hansen, Zhen Song, Andrew Poulson, Anna Komarova, Antoine Baggett, Theodore Kerr, Aleksandra Mikus, Carolina Ortiz Dominguez, Alexandra Tyukavina, and André Lima. Rapid monitoring of global land change. Nature Communications, 16(1), 2025.DOI
- [52]Colorado J. Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uyttendaele, and Trevor Darrell. Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In ICCV, 2023.arxiv.org/abs/2212.14532
- [53]Christian Requena-Mesa, Vitus Benson, Markus Reichstein, Jakob Runge, and Joachim Denzler. EarthNet2021: A large-scale dataset and challenge for earth surface forecasting as a guided video prediction task. In CVPRW, pages 1132–1142, 2021.arxiv.org/abs/2104.10066
- [54]Shawn Ryan, Megan Powell, Joanne Ling, and Li Wen. Streamlining wetland vegetation mapping with AlphaEarth embeddings: Comparable accuracy to traditional methods with cleaner maps and minimal preprocessing. Remote Sensing, 18(2):293, 2026.DOI
- [55]David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3):1181–1191, 2020.arxiv.org/abs/1704.04110
- [56]Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020.arxiv.org/abs/2006.09661
- [57]Michael J. Smith, Luke Fleming, and James E. Geach. EarthPT: A time series foundation model for Earth observation. NeurIPS Climate Change AI Workshop, 2023.arxiv.org/abs/2309.07207
- [58]Pedro Sousa, Will Tebbutt, Sadiq Jaffer, Robin Young, Anil Madhavapeddy, and Richard E. Turner. Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling. arXiv preprint arXiv:2608.12271, 2026.arxiv.org/abs/2608.12271
- [59]Adam J. Stewart, Heng Fang, Isaac A. Corley, and Xiao Xiang Zhu. Earth embeddings. arXiv preprint arXiv:2608.03410, 2026.arxiv.org/abs/2608.03410
- [60]Yuli Sun, Lin Lei, Xiao Li, Xiang Tan, and Gangyao Kuang. Structure consistency-based graph for unsupervised change detection with homogeneous and heterogeneous remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1–21, 2021.DOI
- [61]Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, et al. Prithvi-EO-2.0: A versatile multi-temporal foundation model for earth observation applications. arXiv preprint arXiv:2412.02732, 2024.arxiv.org/abs/2412.02732
- [62]Kyriacos Themistocleous, Christiana Papoutsa, Silas Michaelides, and Diofantos Hadjimitsis. Investigating detection of floating plastic litter from space using Sentinel-2 imagery. Remote Sensing, 12:2648, 2020.DOI
- [63]Gabriel Tseng, Anthony Fuller, Marlena Reil, Henry Herzog, Patrick Beukema, Favyen Bastani, James R. Green, Evan Shelhยamer, Hannah Kerner, and David Rolnick. Galileo: Learning global & local features of many remote sensing modalities. arXiv preprint arXiv:2502.09356, 2025.arxiv.org/abs/2502.09356
- [64]Jan Verbesselt, Rob Hyndman, Glenn Newnham, and Darius Culvenor. Detecting trend and seasonal changes in satellite image time series. Remote Sensing of Environment, 114(1):106–115, 2010.DOI
- [65]Xiaobiao Wang, Shunping Xie, Xueliang Zhang, Cheng Chen, Hao Guo, Jinkang Du, and Zheng Duan. A robust multi-band water index (MBWI) for automated extraction of surface water from Landsat 8 OLI imagery. International Journal of Applied Earth Observation and Geoinformation, 68:73–91, 2018.DOI
- [66]Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M. Albrecht, and Xiao Xiang Zhu. SSL4EO-S12: A large-scale multimodal, multitemporal dataset for self-supervised learning in Earth observation. IEEE Geosci. Remote Sens. Mag., 11(3):98–106, 2023.
- [67]Haoran Wei, Kun Jia, Qiao Wang, Biao Cao, Jianbo Qi, Wenwu Zhao, and Jian Yang. Real-time remote sensing detection framework of the earth’s surface anomalies based on a priori knowledge base. International Journal of Applied Earth Observation and Geoinformation, 122, 2023. 7, 10
- [68]Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. ETSformer: Exponential smoothing transformers for time-series forecasting. arXiv preprint arXiv:2202.01381, 2022. 3, 7, 9arxiv.org/abs/2202.01381
- [69]Zhangyu Wu, Gengchen Mai, Ni Lao, and Krzysztof Janowicz. LocDiff: Identifying locations on Earth by diffusing in the Hilbert space. In NeurIPS, 2025. 12arxiv.org/abs/2503.18142
- [70]Haiping Xia, Yunhao Chen, and Jinling Quan. A simple method based on the thermal anomaly index to detect industrial heat sources. International Journal of Applied Earth Observation and Geoinformation, 73:627–637, 2018. 10DOI
- [71]Zhitong Xiong, Yi Wang, Fanghong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired multimodal foundation model for Earth observation. arXiv preprint arXiv:2403.15536, 2024. 3, 13
- [72]Hanqiu Xu. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing, 27:3025–3033, 2006. 10DOI
- [73]Yifang Yin, Zhenguang Liu, Ying Zhang, Sheng Wang, Rajiv Ratn Shah, and Roger Zimmermann. GPS2Vec: Towards generating worldwide GPS embeddings. In ACM SIGSPATIAL, 2019. 13
- [74]Yuxiang Zhang, Shunlin Liang, Wenyuan Li, Han Ma, Jianglei Xu, Yichuan Ma, Jiangwei Xie, Wei Li, Mengmeng Zhang, Ran Tao, and Xiang-Gen Xia. UniTS: Unified spatio-temporal generative model for remote sensing. arXiv preprint arXiv:2512.04461, 2025. 1, 3arxiv.org/abs/2512.04461
- [75]Zhe Zhu and Curtis E. Woodcock. Continuous change detection and classification of land cover using all available Landsat data. Remote Sensing of Environment, 144:152–171, 2014. 3
- [76]Zhe Zhu, Junxue Zhang, Zhiqiang Yang, Amal H. Aljaddani, Warren B. Cohen, Shi Qiu, and Congliang Zhou. Continuous monitoring of land disturbance based on Landsat time series. Remote Sensing of Environment, 238:111116, 2020. 7, 10DOI

