betaFeedback

WorksComputer Science

Global Solar Siting from Geospatial Embeddings

Abstract

Global solar photovoltaic (PV) capacity is rapidly growing, helping decrease our emissions and dependence on fossil fuels. However, optimally siting a solar plant remains challenging, requiring anticipation of geographical, urban, and infrastructure needs. Current methods rely on multi-criteria decision analysis (MCDA) using hand-picked geophysical layers and solar-resource products, however these strategies are not geographically transferable and do not capture the practical requirements of solar sites. We use geospatial embeddings, dense vectors representing geophysical properties of a location, to analyse 38,591 existing solar sites and the land they were built on to learn the latent solar suitability signal. We assess a set of linear and non-linear models and select a soft-voting ensemble of a random forest and a multilayer perceptron (RF+MLP ensemble), chosen for accuracy and large scale deployability. Validated against future installations in thirteen countries, the ensemble separates existing solar sites from random land with a median ROC-AUC of 0.90. Compared with published MCDA strategies, the ensemble increases ROC-AUC by 0.06-0.30 in every country. The ranking also concentrates real deployment: the most-suitable 20% of land holds 84% of estimated generation (annual energy yield) and 67% of installed capacity (potential power output). After applying standard siting constraints, 5.6% of global land (9.8% if cropland is permitted) remains highly suited to solar. The global 10 m resolution solar suitability dataset and code can be found here: https://github.com/Abdullahr29/Global-Solar-Siting.

Introduction

Energy production and consumption accounts for approximately 76% of global greenhouse gas emissions [1], primarily due to the burning and consumption of fossil fuels [2]. Meeting net-zero targets requires renewable energy sources to displace fossil fuel driven generation at an unprecedented pace [3]. Solar photovoltaic (PV) is a key technology in this energy transition, accounting for 81% of all new renewable capacity added in 2024, and reaching 3 terawatt (TW) in installed capacity globally in 2026 [4]. By 2030, the world is projected to be installing over 1 TW of solar annually, and as deployment accelerates and highly accessible locations are developed, the binding constraint is increasingly in where to build and identifying land that will make a successful site [4, 5].

Solar siting currently requires the analysis of several geospatial features pertaining to topography, land cover, solar resources, and urban infrastructure [6]. These layers are used in Multi-Criteria Decision Analysis (MCDA) frameworks to optimise solar placement, but these approaches are highly variable, with decisions driven by feature weighting and often limited to one country or region [7, 8, 9, 10, 11].

Machine learning (ML) approaches in solar research have primarily focused on forecasting solar energy generation [12, 13]. Studies which apply ML to solar siting follow a similar pattern to MCDA, engineering geospatial features for classical models, with decision tree ensembles often performing well but still being limited to one region [14, 15, 16, 17].

Where global products do exist they model PV power potential [18, 19] or map existing installations, such as the Global Solar Power Tracker [20]. Solar siting reflects not only solar-resource availability but also policy, infrastructure and land-use constraints [21]. A forward-looking global solar suitability layer would supplement these products and support faster, better-justified siting.

Learning the inherent property of land which is suitable for solar through a set of predefined features is challenging. Recent advancements in geospatial representation learning offer a new pathway to tackling this problem [22]. Global embeddings such as AlphaEarth Foundations (AEF) encode a diverse mix of geophysical and meteorological properties alongside remote sensing information into a high dimensional vector [23], and even simpler models trained on embeddings can transfer land-cover labels across continents [24]. While geospatial embeddings are not as interpretable as data from optical sensors such as Sentinel-2, they circumvent many of the challenges that come with using raw imagery, including cloud masking, seasonal adjustments, and processing differences, challenges faced in solar infrastructure detection [25, 21]. Embeddings instead provide an analysis-ready, gap-filled annual summary of every 10 m pixel, learnt from many sensors at once [23].

Aims and Objectives. In this study, we explore whether geospatial embeddings can encode a non-obvious, future-looking property of land: the probability and suitability of future solar development. We evaluate a range of models trained on geospatial embeddings of existing solar sites and non-solar land, validating their performance on future solar installations across countries, and comparing them against published MCDA approaches and existing solar-resource products. Finally, we release a global 10 m solar suitability layer, empowering future solar siting and the renewable energy transition.

Data

Positive sites. From the Global PV Facility (GPVF) inventory of 140 945 facilities (with delineated footprints and installation dates) we draw a country-balanced sample of 38 591 sites [26]. 64-dimensional AEF embeddings of the sites 2 years before solar installation are then used for training; the AEF embedding record begins in 2017, therefore we can only study sites from 2019 to 2024. The sample spans 154 countries. Each site’s delineated footprint is rasterised onto the 10 m embedding grid and the pixels inside it form the positive labels. The task is to anticipate development rather than to recognise existing panels, as shown in Figure 1.

Figure 1. The pipeline for learning solar suitability. For a known solar site, the per-pixel geospatial embedding of the land (here the 64-dimensional AEF embedding) is sampled two years before installation, capturing the pre-installation conditions. A random forest and multilayer perceptron ensemble, trained on solar and globally-random non-solar land, maps each embedding pixel to P(solar)P(\mathrm{solar}). Post-installation, the true solar asset footprint (outlined) shows the developed site. Sentinel-2 imagery is shown for visual reference only and is not a model input. Applied globally, the model yields an openly-released global 10 m suitability layer.

Negative sites. Non-solar land is sampled as globally-random parcels at least 5 km from any solar site of any year, present in either the GPVF or the TZ-SAM, a secondary global inventory of solar energy facilities [27], datasets. AEF embeddings for negative sites are sampled randomly from any year between 2017-2022. The final solar-to-non-solar ratio is ∼1.5\sim1.5.

Testing & validation. Scenes are split 80/20 into train and test as whole chips, to minimise pixel leakage (see Appendix § 7 for exact values). We additionally evaluate the models against the TZ-SAM inventory, testing specifically against sites missing from the GPVF dataset, our training database.

Methodology

Model benchmarking. To assess the optimal mechanism of extracting solar suitability from the AEF embedding, we compare nine models of increasing complexity: a naive mean-embedding prototype scored by cosine similarity [28]; a quadratic discriminant (QDA) [29]; a logistic probe [30]; MaxEnt, the standard species-distribution and GIS-suitability baseline [31]; random Fourier features with a logistic head [32]; a random forest (RF) [33]; histogram gradient boosting [34]; a multilayer perceptron (MLP) [35]; and a soft-voting ensemble of the RF and MLP (RF+MLP ensemble) [36]. All nine are trained on a common balanced sample of training pixels; full formulations and hyperparameters are given in Appendix § 8.

Validation. On held-out pixels we report the area under the Receiver Operating Characteristic curve (ROC-AUC), the probability that a randomly chosen solar pixel is ranked above a randomly chosen non-solar pixel, and the area under the precision–recall curve (PR-AUC, equivalently average precision), which is more informative under the ∼\sim1:5 class imbalance. We further validate the RF+MLP ensemble forward in time, across 13 held-out countries and under a leak-proof temporal split, cross-checking every country against the independent TZ-SAM inventory; the full protocol, chip-level and per-country results, and leakage ablations are given in Appendix § 9 and Appendix § 10.

Global-scale analysis. We investigate what the suitability ranking implies for real deployment. Joining each validation site to its TZ-SAM capacity estimate, we measure how the world’s installed capacity distributes across the ranked land. Joining instead to the GPVF modelled annual generation, we ask whether the ranking concentrates estimated generation, testing the correlation between suitability and energy yield. Finally, we threshold the suitability map at the tenth percentile of real-site scores, and impose literature-derived siting cut-offs (slope ≤5∘\le5^\circ, grid distance ≤10\le10 km, protected-area exclusion) to report a grid-constrained realistic fraction of global land highly suited for solar installation.

Comparison to MCDA. We benchmark the ensemble model against the MCDA that underpins conventional siting. Using two sets of published expert weights (Richards et al. [10] and Chen Jong and Mohamud Ahmed [11]) and two standard aggregators (weighted linear combination and TOPSIS), we construct suitability maps from solar-relevant features and score each MCDA approach on the identical site-versus-random task against the ensemble. MCDA model explanations, weights, and parameters are given in Appendix § 14. Finally, we conduct a covariate analysis to quantify the extent to which the ensemble’s suitability signal is explained by standard GIS layers, including land cover, slope, irradiance, and solar resource products such as PVOUT (see Appendix § 13).

Results

The siting signal is real but strongly non-linear. Table 1 ranks the extractors on held-out pixels and on held-out sites per country. The naive mean-embedding prototype—the classic embedding similarity search mechanism—separates solar from non-solar land only weakly (ROC-AUC 0.70), so suitability is not a simple direction in embedding space. Linear and kernel models recover most of the signal (logistic and MaxEnt 0.87, random Fourier features 0.89) and the non-linear models reach 0.92. Averaging a random forest and an MLP trained on the same sample improves on every single model across every validation test, and this ensemble is therefore chosen for deployment. The released global layer is the same RF+MLP ensemble refit on all data.

MethodFamilyPixel ROC-AUCPixel PR-AUCCountry ROC-AUCTZ-SAM ROC-AUC
Mean-embedding cosineprototype0.7020.3180.7510.739
QDAdiscriminant0.8540.5180.8300.834
Logistic probelinear0.8660.5510.8280.823
MaxEntlog-linear (SDM)0.8740.5770.8330.827
Random Fourier featureskernel0.8910.6280.8560.841
Random foresttree ensemble0.9140.6960.8930.874
Hist. gradient boostingboosted trees0.9190.7130.8960.880
MLPneural0.9240.7280.8930.872
RF+MLP ensembleensemble0.9320.7520.9040.882

Table 1. Extraction-method benchmark. Pixel: held-out test pixels (PR-AUC chance rate 0.17). Country: ROC-AUC of future sites against random national land, averaged over the thirteen validation countries weighted by site count. TZ-SAM: independent sites absent from our training inventory. All methods are trained on one common balanced sample; combining a random forest and an MLP into an ensemble improves on every single model in the benchmark.

The model anticipates future solar. Trained only on land as it appeared at least two years before installation, the model separates real future sites from random national land in all thirteen countries, with forward ROC-AUC spanning 0.83–0.95 (see Appendix § 10). As Figure 2 demonstrates: at site scale the model scores the pre-installation embedding of land that later became solar as highly suitable (Figure 2a), and at national scale (Figure 2b, Greece) the 876 validation sites concentrate on the suitable minority, their scores cleanly separated from random land (inset histogram, ROC-AUC 0.894).

Figure 2. (a) A solar site (Japan) with footprints outlined: Sentinel-2 before construction (2017), the AEF embedding two years before installation (2017), Sentinel-2 after construction (2020), and the per-pixel RF+MLP suitability map; Sentinel-2 is shown for visual reference only. The land is forested hills and valleys in 2017, yet the pre-installation embedding scores them accurately. (b) National suitability for Greece (AEF-2019) with 876 validation sites concentrated on the most-suitable land: the score histogram shows them cleanly separated from random national land (ROC-AUC 0.894). (c) Cumulative share of solar sites, installed capacity, and estimated generation captured versus the most-suitable land ranked by the ensemble; capacity concentrates less than sites and generation, as the largest farms are atypically sited (see Appendix § 12).

The map concentrates capacity and generation. The most-suitable 20% of land carries 67% of the world’s installed capacity and 84% of its estimated generation (Figure 2c), and weighting validation sites by their generation lowers discrimination only slightly (ROC-AUC 0.93 to 0.90), as the largest plants are less typically sited. Globally, 22.1% of land clears the suitability threshold and 5.6% (9.8% if cropland is allowed) survives the grid-and-terrain constraints, so suitable land is not what limits solar, economics and policy are; the map is best used to prioritise within that abundant land.

The model surpasses existing siting methods. The ensemble outperforms published MCDA in every one of the thirteen countries, by 0.06–0.30 ROC-AUC, every gap statistically significant (Appendix § 14, Table 5). Figure 3 shows China, where MCDA ranks the empty western interior suitable, yet real solar is built in the east, which the ensemble correctly identifies. Regressing the ensemble’s score on standard GIS layers (Appendix § 13), land cover and slope are the most important predictors in most countries, while irradiance and the long-term PV output potential (PVOUT) rarely are. PVOUT alone cannot separate sites from random land (ROC-AUC 0.49). Between 20% and 53% of the ensemble score’s variance is not explained by these geospatial layers, information the embedding carries beyond hand-picked features.

Three maps of China comparing RF+MLP suitability, published MCDA, and their disagreement

Figure 3. Both maps are shown as within-country percentiles (the share of China’s land each cell outscores). (a) RF+MLP suitability (AEF-2019) with future solar sites (blue; a subsample of 1500 of the 27 158 scored sites): solar is built in the east, which the ensemble ranks high (ROC-AUC 0.91). (b) Published MCDA instead follows infrastructure and solar-resource heuristics, ranking western and northern land suitable (ROC-AUC 0.61, the best of the 16 tested MCDA configurations). (c) The difference between the RF+MLP ensemble and MCDA. Red indicates a high ensemble and low MCDA score (blue vice-versa), outlining significant disagreement in high solar regions.

CountrynnRF+MLPMCDAGap [95% CI]
China27 1580.9140.615+0.299 [0.291, 0.307]
Chile5440.9610.725+0.237 [0.218, 0.256]
Greece17350.9160.683+0.233 [0.218, 0.249]
United States75640.9290.741+0.189 [0.180, 0.197]
India43860.8860.712+0.174 [0.165, 0.184]
Poland20620.8450.671+0.174 [0.161, 0.187]
Spain27750.8940.720+0.174 [0.163, 0.186]
Japan32460.9410.779+0.161 [0.151, 0.172]
Malaysia1630.9370.788+0.149 [0.107, 0.197]
Germany28640.8860.740+0.146 [0.135, 0.157]
Philippines1670.9280.831+0.097 [0.070, 0.125]
South Africa2700.9210.828+0.093 [0.070, 0.116]
Colombia1980.9630.905+0.058 [0.041, 0.078]

Table 5. RF+MLP ensemble versus published MCDA, combined GPVF+TZ-SAM sites, ordered by gap. MCDA is the best of sixteen configurations (Richards/Chen weights and their equal-weight variants × WLC/TOPSIS × exclusion/weighted land). Bracketed ranges are paired bootstrap 95% confidence intervals on the gap; all lie strictly above zero.

Discussion

Our work shows that geospatial embeddings can be probed for non-obvious, forward-looking properties of land, and that solar suitability is a strongly non-linear but latent property, best extracted by combining complementary models. Future extensions of this work include the investigation of other geospatial embedding products [37] and spatial models which look beyond individual pixels and consider site suitability across scales.

Our current approach is built on the assumption that existing solar facilities are well sited, however this may limit the model’s ability to identify land that is suitable in underdeveloped regions. Future approaches applying positive-unlabelled learning may mitigate this limitation, by treating unlabelled land as a mixture of suitable and unsuitable rather than as negatives. Conversely, land which has been identified as highly suitable may not remain so in the future due to climate change. Risks to solar power are emerging, to solar irradiance (wildfire smoke) [38] or to solar assets themselves (flooding) [39], which the current model cannot foresee. This necessitates the development of future solar suitability models which incorporate climate change projections and risk assessment.

Access to solar power presents an opportunity for clean and affordable energy in developed and developing nations alike. An open, global and validated suitability layer can help governments, grid and land-use planners, and development organisations prioritise where to investigate, including in data-poor countries. Used in conjunction with environmental assessment and community consultation, the suitability layer can serve as an additional resource for data-informed siting of future solar plants with minimal environmental and social impact.

Acknowledgments and Disclosure of Funding

This work was funded by the European Space Agency (ESA) Dragon Cooperation Programme through the project FUCEO: Exploring synergies between Chinese and European EO mission using data fusion (funding ID 95452). This work used JASMIN, the UK’s collaborative data analysis environment (https://www.jasmin.ac.uk), and Google Earth Engine for data access and processing. The authors thank the ESA Φ–lab and ESA Climate Office for their support during this research.

Training Data Size

Table 2 summarises the training data. Positive pixels are the rasterised footprints of the country-balanced solar-site sample; negative pixels are drawn from the globally-random non-solar parcels (Section 2). The number of training pixels and positive-to-negative ratio were tested against model performance, with the current ∼\sim1:5 solar-to-non-solar pixel ratio performing optimally. The 80/20 train/test split is made over whole scenes, so no two chips from the same site fall on opposite sides of the split.

ClassSourceChips (train : test)Pixels (train : test)
Positive (solar)PV-facility inventory; footprint masks65 811 : 16 3376 193 548 : 1 550 685
Negative (non-solar)Random parcels ≥\geq5 km from any solar site; TZ-SAM-screened10 775 : 2 66530 170 000 : 7 462 000
Total76 586 : 19 00236 363 548 : 9 012 685

Table 2. Distribution of training data by chips and pixels. Scenes are split 80/20 into train and test as whole chips.

Extraction methods: formulations and parameters

All methods take the 64-dimensional AEF pixel embedding x∈R64\mathbf{x} \in\mathbb{R}^{64} as input. The benchmark models (Table 1) are trained on a common balanced subsample of 1 400 000 training pixels. Models that benefit from standardisation are fitted on features centred and scaled by a StandardScaler learnt on the subsample. We use the scikit-learn implementations [40] (unlisted arguments take library defaults), and PyTorch for the tuned MLP.

Mean-embedding prototype (cosine). A prototype μ\mu is the mean of up to 2 000 000 (no change in vector with more training pixels) solar-pixel embeddings, unit-normalised to μ^\hat{\mu}; a pixel is scored by s(x)=xTμ^s(\mathbf{x}) = \mathbf{x}^{\mathsf{T}}\hat{\mu}. This is the classic embedding-similarity approach and involves no discriminative training. Because AEF embeddings are themselves near unit-norm, normalising the prototype (not the inputs) recovers cosine similarity.

Logistic probe. ℓ2\ell_2-regularised logistic regression (C=1C = 1, max_iter = 300) on standardised embeddings with balanced class weights, the simplest discriminative linear read-out.

Quadratic discriminant analysis (QDA). Class-conditional Gaussians with separate covariance matrices, yielding a quadratic decision boundary; fitted on the raw embeddings.

MaxEnt. The maximum-entropy species-distribution model [31], the standard ecological/GIS suitability baseline. Where the elapid implementation is available we use it directly; otherwise we use its infinitely-weighted-logistic-regression equivalent: an ℓ1\ell_1-penalised logistic regression (C=0.5C = 0.5, saga solver, balanced weights) on linear and quadratic embedding features [x,x2][\mathbf{x}, \mathbf{x}^2].

Random Fourier features (RFF). A 512-dimensional random Fourier feature map [32] approximating an RBF kernel (γ=1/64\gamma= 1/64), followed by balanced logistic regression.

Random forest. Benchmark: 100 trees, max_depth = 16, min_samples_leaf = 50, balanced class weights, on the common subsample. Tuned (ensemble member): 80 trees, max_features = 12, min_samples_leaf = 14, full bootstrap samples, Gini criterion, balanced class weights, on the full training split, chosen by a sweep over feature subsampling, leaf size and tree count without any size limit.

Histogram gradient boosting. Gradient-boosted decision trees (max_iter = 300, learning rate 0.1, balanced weights)—the most accurate single tree model on the common subsample.

MLP. Benchmark: two hidden layers (128→\rightarrow64), ReLU activations and early stopping, trained on standardised embeddings. Tuned (ensemble member): two hidden layers of 256 ReLU units, no dropout, standardised inputs, class-balanced binary cross-entropy, AdamW (learning rate 3×10−33 \times10^{-3}, weight decay 10−510^{-5}), batch size 16 384 and a one-cycle learning-rate schedule over eight epochs, stopped after the first epoch (the epoch chosen on the selection half); the prediction is the mean of five seeds. The configuration was selected by a random search over 64 configurations [41] (depth and width, dropout, weight decay, learning rate, batch size, schedule length and warm-up, input standardisation), scoring every configuration as the RF’s partner in the ensemble.

RF+MLP ensemble (selected). Soft voting: P(solar∣x)=12PRF(x)+12PMLP(x)P(\mathrm{solar}\mid x) = \frac{1}{2}P_{\mathrm{RF}}(x) + \frac{1}{2}P_{\mathrm{MLP}}(x) [36]. All selection decisions (MLP configuration, stopping epoch, finalist, ensemble weight) were made on one half of the country validation sites, split by 0.5∘0.5^\circ cells, under a rule fixed before the confirmation results were seen: ties within 0.001 ROC-AUC go to the cheaper model, and the weight moves from 12\frac{1}{2} only for a gain of at least 0.001 (none did). Learned stacking and rank averaging were tested, but performed identically.

Additional site-scale examples

Figure 4 complements the single site of Figure 2a with three further installations spanning an urban coastal fringe (Australia), an agricultural interior (United States) and a tropical smallholder landscape (Philippines). For each we show the Sentinel-2 scene before and after construction, the AEF embedding two years before installation, and the per-pixel RF+MLP suitability on both a linear scale and a nonlinear (γ=2.2\gamma= 2.2) rescaling that expands the high-suitability end of the range on an absolute 0 to 1 axis. In every case the pre-installation embedding of the land under the footprint (cyan) is scored as suitable; the uniformly-bright Philippines map illustrates the need for nonlinear rescaling, an issue also discussed in Appendix § 11 for dim countries.

Site-scale suitability maps for three further installations

Figure 4. Site-scale suitability for three further installations (rows: Australia, United States, Philippines). Columns: Sentinel-2 before and after construction, the AEF embedding two years before installation, and the RF+MLP suitability map on a linear and a nonlinear (γ=2.2\gamma= 2.2) colour scale; footprints outlined in cyan. Across all three geographies the embedding scores the future-solar parcels as suitable before any panels appear, and the nonlinear scale highlights the high-suitability cores.

Forward validation

Metrics and percentile capture. Beyond ROC-AUC and PR-AUC, we summarise deployment behaviour with a percentile-capture test: each real site is assigned the fraction of random national land it outscores, its within-country suitability percentile, from which we report the share of sites in the most-suitable 20%, 10% and 5% of land, over the combined GPVF and TZ-SAM validation set.

A typical future site sits at the 95th percentile of national land, with 85%, 66% and 48% of sites in the most-suitable 20%, 10% and 5% of land respectively.

Per-country results. The ensemble separates real future sites from random national land in all thirteen countries (Table 3, Figure 5). The inventory and TZ-SAM sources agree closely (median ∣ΔROC∣≈0.008|\Delta\mathrm{ROC}| \approx0.008), and pooling the far larger TZ-SAM positive counts tightens the intervals for the data-poor countries by a median factor of 1.7. Countries are ordered by combined ROC-AUC.

Table 3: Per-country forward validation of the RF+MLP ensemble, against inventory and independent TZ-SAM future sites. nn gives the number of validation sites; each is scored against random national land points. Bracketed ranges are bootstrap 95% confidence intervals.

InventoryTZ-SAMCombined
CountrynnROC-AUC [95% CI]nnROC-AUCnnROC-AUC [95% CI]
CHL810.950 [0.938, 0.962]1630.9552440.953 [0.941, 0.964]
COL230.969 [0.955, 0.982]930.9481160.952 [0.939, 0.963]
JPN14500.935 [0.929, 0.941]5170.92919670.934 [0.928, 0.939]
PHL220.930 [0.893, 0.964]510.924730.926 [0.903, 0.946]
USA29930.936 [0.931, 0.941]22000.90351930.922 [0.917, 0.927]
CHN29780.924 [0.918, 0.929]11 0840.90314 0620.907 [0.902, 0.912]
MYS320.853 [0.801, 0.898]480.932800.900 [0.871, 0.929]
GRC2210.894 [0.880, 0.906]6550.8948760.894 [0.885, 0.901]
ZAF280.899 [0.819, 0.965]1400.8911680.893 [0.864, 0.918]
ESP3970.867 [0.853, 0.879]8170.86412140.865 [0.856, 0.874]
IND4640.868 [0.853, 0.884]20860.86025500.862 [0.854, 0.870]
DEU4650.885 [0.871, 0.899]9520.84614170.858 [0.849, 0.868]
POL1900.853 [0.832, 0.872]10420.82112320.826 [0.816, 0.837]

Table 3.

Figure 5. Per-country forward ROC curves for the RF+MLP ensemble on inventory sites installed from 2021 onward, scored against random national land. Step-like curves reflect the small validation counts of the data-poor countries (Table 3).

Further validation. We validate the ensemble along three axes, each isolating a bias source. (i) Forward in time: each national map uses the 2019 embedding while validation sites were installed from 2021 onward, and any validation site belonging to a training-split site is removed. *(ii) Across space (leave-one-country-out, LOCO): the target country is withheld from training, testing transfer to unseen geographies. (iii) Leak-proof temporal: training is restricted to imagery up to 2021 (installs up to 2023) across all countries and validated on 2024 installs, on the 2022 embedding. Every country is additionally cross-checked against the independent TZ-SAM inventory, a different source absent from training, with bootstrap 95% confidence intervals. The LOCO and temporal models are deliberately handicapped diagnostics (the same RF+MLP recipe refit without the withheld country or years) that bound the two leakage pathways; all other results use the ensemble trained on the 80% training split, and the released layer uses the same recipe refit on all data.

Figure 6 outlines both LOCO and temporal ablation results. For the spatial holdout (nine countries), performance falls furthest for the hardest, most distinctive countries (China 0.92 → 0.72, Poland 0.85→0.74, India 0.87→0.78) and least for the easier ones (Spain, Philippines, Chile), averaging 0.82, so the optimism premium tracks country difficulty rather than being uniform. The temporal hold-out (twelve countries) is fully leak-proof in time; it stays strong (mean 0.90, pooled global-2024 0.89) and matches or exceeds the in-sample forward number in six of the ten countries with both (e.g. India, Spain, Germany), confirming the forward result is not a temporal-leakage artefact.

Figure 6. Site-versus-random ROC-AUC of the RF+MLP ensemble (full model) against the same recipe refit as hold-out models, per country. Left: leave-one-country-out spatial hold-out. Right: leak-proof temporal hold-out (training imagery ≤2021\le2021, 2024 installs; countries with a forward set).

Weak-country analysis

Performance ranges from ROC-AUC 0.83 (Poland) to 0.95 (Chile), but this spread reflects country difficulty, not training-data volume: per-country ROC correlates weakly negatively with the number of sites (Spearman ρ=−0.26\rho= -0.26; China, with 14 000+ sites, scores 0.91; Chile, with 244, scores 0.95), so the harder countries are geophysically diverse rather than data-starved. The effect is one of readability, not skill. In a dim country such as India the raw scores are compressed into a narrow band (Fig. 7a), so re-expressing each country’s land as percentiles of its own score distribution brightens the map (Fig. 7b) while leaving ROC-AUC unchanged at 0.868. This makes the percentile view more readable for national maps. The raw scores nonetheless share one global scale and each country’s future sites outrank the random land of the other twelve countries (ROC-AUC 0.83–0.96), …

Figure 7. (a) The RF+MLP suitability field on an absolute 0 to 1 scale: India’s land scores stay low, so the country map looks dim. (b) Re-expressing the same field as within-country percentiles brightens it and reveals structure; overlaid validation sites (blue) concentrate in the high-percentile regions. (c) Score histogram for random land versus solar sites (ROC-AUC 0.868); the percentile transform is rank-preserving, so it leaves ROC-AUC unchanged.

Discrimination by installed capacity

We bin the combined GPVF+TZ-SAM validation sites by their TZ-SAM installed-capacity estimate and compute the site-versus-random ROC-AUC within each capacity class (Figure 8). Discrimination is strong across the small-to-mid solar farms that dominate deployment (ROC-AUC 0.90 below 5 MW) and falls for the rarer, atypically-sited utility farms above 100 MW (0.79), whose siting is often driven by land availability and heavy transmission infrastructure rather than the qualities the model has learnt. Because the model scores pixels, it does not test whether enough contiguous suitable land exists for a plant of a given size.

Figure 8. Site-versus-random ROC-AUC of the RF+MLP ensemble by installed capacity (bootstrap 95% CIs): strong for the small-to-mid sites that dominate deployment, declining for the rare >100 MW utility farms.

Covariate analysis

Method. To test whether the embedding carries siting information beyond standard GIS layers, we build a covariate stack for each country: land cover (ESA WorldCover 2021 [42]), slope and elevation (Copernicus GLO-30 DEM [43]), air temperature, the irradiance components GHI (global horizontal irradiance), GTI (global tilted irradiance), DNI (direct normal irradiance) and DIF (diffuse horizontal irradiance), and long-term PVOUT (photovoltaic output), all from the Global Solar Atlas [44]; distance to roads (OpenStreetMap [45]) and to the electricity grid (Gridfinder [46]); population density (GPWv4.11 [47]); and protected-area status (WDPA [48]). We regress the ensemble’s suitability score on the covariate stack with both a linear model and a random forest. The residual that the covariate model cannot reconstruct (1−R21 - R^2) measures the information the embedding holds that these layers do not; per-covariate permutation importance reveals which physical factors the learned suitability aligns with. Figure 9 shows part of this stack for the United States.

Figure 9. The physical-covariate stack for the United States alongside our RF+MLP suitability (top-left, AEF-2019, raw P(solar)P(\mathrm{solar})). Suitability is high across the developed central and eastern land and low over the remote mountain west, tracking land cover, terrain and infrastructure (grid- and road-distance) rather than the irradiance layers (GHI, PVOUT), which peak in the south-west where comparatively little solar is actually built.

Result. The residual runs from 0.20 (Chile) to 0.53 (India), and the random-forest fit consistently exceeds the linear one, confirming the information is non-linear. The factors the suitability aligns with most are land cover and terrain (slope, elevation) and air temperature. Infrastructure and irradiance (GHI/GTI/PVOUT) matter only in a few resource-driven countries and are rarely dominant (Figure 10).

Figure 10. Per-country covariate permutation importance of a model predicting the RF+MLP ensemble suitability score from GIS layers, with the embedding-novelty residual under each country (higher = more the embedding knows that the layers do not). WorldCover classes are summed into a single land-cover row.

Orthogonality to solar resource products. The learned suitability is essentially orthogonal to the authoritative long-term photovoltaic output potential (PVOUT) product [18]: across countries their rank correlation is near zero or weakly negative (Spearman ρ≈−0.10\rho\approx-0.10), and PVOUT alone cannot separate real sites from random land (ROC-AUC 0.49 versus our 0.90). Conventional siting anchored on such resource layers therefore misranks deployment, as the China comparison (Figure 3) shows.

MCDA benchmark

MCDA construction. We reconstruct multi-criteria suitability from two published expert weightings: an infrastructure-led scheme [10] and a resource-and-terrain scheme [11], listed in Table 4. Criteria are built from the same data layers as the covariate analysis (Appendix § 13), with aspect derived from the DEM. Each criterion is scaled to [0, 1] by country-relative 2–98% min–max normalisation.

CriterionRichards et al.Chen et al.
Road distance0.54–
Grid distance0.140.17
Air temperature0.080.17
Aspect0.08–
Slope0.060.16
GHI0.040.19
PVOUT0.03–
Elevation0.020.16
GTI0.02–
Population density–0.15

Table 4. Published expert criterion weights reconstructed for the MCDA benchmark, from Richards et al. [10] and Chen Jong and Mohamud Ahmed [11]; a dash marks a criterion the scheme omits and each column sums to one. GHI/GTI: global horizontal/tilted irradiance; PVOUT: long-term photovoltaic output potential.

Aggregation and land constraints. We combine the weighted criteria with two standard aggregators. Weighted linear combination (WLC) is the weighted sum of the normalised criteria. TOPSIS

(Technique for Order of Preference by Similarity to Ideal Solution) instead ranks each location by its relative closeness to a fixed ideal point that is best on every criterion (all ones) against an anti-ideal that is worst on every criterion (all zeros): the score is d−/(d++d−)d^{-}/(d^{+}+d^{-}), with d+d^{+} and d−d^{-} the Euclidean distances to the ideal and anti-ideal, so a location ranks high when it lies near the ideal and far from the anti-ideal. Land cover and protected areas enter in one of two ways: either as a hard exclusion only—WDPA protected areas and unsuitable WorldCover classes (open water, built-up, permanent ice) forced to zero—or, additionally, as a continuous weighted land-suitability criterion layered on top of that exclusion. Crossing the two weight sets, the two aggregators, and the two land treatments yields eight published-weight MCDA configurations per country; we add their equal-weight counterparts (each scheme’s criteria weighted uniformly), giving sixteen configurations. Each is scored on the same combined GPVF+TZ-SAM site-versus-random task as the ensemble (paired bootstrap over sites), and we report the strongest, a deliberately favourable baseline for MCDA.

Result. The ensemble beats the best MCDA configuration in all thirteen countries, and every gap is significant (Table 5). The advantage is largest exactly where expert weights misread the geography, China (+0.30), where MCDA over-weights the irradiance that anti-correlates with deployment, and smallest where siting genuinely tracks the resource: Colombia, South Africa). Figure 11 shows the Colombia case (the analog of Figure 3): a solar resource driven country where the two largely agree.

Colombia RF+MLP suitability map

(a)

Colombia published MCDA suitability map

(b)

Colombia RF+MLP minus MCDA disagreement map

(c) Colombia RF+MLP-versus-MCDA maps, the analog of Figure 3 for a resource-driven country. (a) RF+MLP suitability (AEF-2019) with the 198 scored validation sites (blue), concentrated in the Andean populated corridor (ROC-AUC 0.96). (b) Best published MCDA (Chen weights, GHI+population; ROC-AUC 0.91). (c) RF+MLP−MCDA disagreement. Unlike China, the ensemble and MCDA largely agree here so the ensemble’s edge is small (+0.058), illustrating that the advantage concentrates where expert weights misread the geography. Certain areas of disagreement, particularly protected areas, are still visible.

Figure 11.

References

  1. [1]Greenhouse Gas (GHG) Emissions Climate Watch. URL https://www.climatewatchdata.org/ghg-emissions.
  2. [2]IEA. Greenhouse Gas Emissions from Energy - Data product. URL https://www.iea.org/data-and-statistics/data-product/greenhouse-gas-emissions-from-energy.
  3. [3]Nadarajah Kannan and Divagar Vakeesan. Solar energy for future world: - A review. Renewable and Sustainable Energy Reviews, 62:1092–1105, September 2016. ISSN 1364-0321. doi: 10.1016/j.rser.2016.05.022. URL https://www.sciencedirect.com/science/article/pii/S1364032116301320.DOI
  4. [4]SolarPower Europe. Global market outlook for solar power 2025-2029, May 2025. URL https://www.solarpowereurope.org/insights/outlooks/global-market-outlook-for-solar-power-2025-2029.
  5. [5]Will Gorman, Julie Mulvaney Kemp, Joseph Rand, Joachim Seel, Ryan Wiser, Nick Manderlink, Fredrich Kahrl, Kevin Porter, and Will Cotton. Grid connection barriers to renewable energy deployment in the United States. Joule, 9(2):101791, February 2025. ISSN 2542-4351. doi: 10.1016/j.joule.2024.11.008. URL https://www.sciencedirect.com/science/article/pii/S2542435124005038.DOI
  6. [6]Abdulaziz Alhammad, Qian (Chayn) Sun, and Yaguang Tao. Optimal Solar Plant Site Identification Using GIS and Remote Sensing: Framework and Case Study. Energies, 15(1), January 2022. ISSN 1996-1073. doi: 10.3390/en15010312. URL https://www.mdpi.com/1996-1073/15/1/312.DOI
  7. [7]Volodymir Kozlov and Wojciech Sałabun. Challenges in reliable solar panel selection using MCDA methods. Procedia Computer Science, 192:4913–4923, January 2021. ISSN 1877-0509. doi: 10.1016/j.procs.2021.09.269. URL https://www.sciencedirect.com/science/article/pii/S1877050921020081.DOI
  8. [8]Vytautas Matulaitis, Gintė Straukaitė, Brian Azzopardi, and Eduardo A. Martinez-Cesena. Multi-criteria decision making for PV deployment on a multinational level. Solar Energy Materials and Solar Cells, 156:122–127, November 2016. ISSN 0927-0248. doi: 10.1016/j.solmat.2016.02.015. URL https://www.sciencedirect.com/science/article/pii/S0927024816000775.DOI
  9. [9]Jack Mathebula and Nhlanhla Mbuli. Application of TOPSIS for Multi-Criteria Decision Analysis (MCDA) in Power Systems: A Systematic Literature Review. Energies, 18(13):3478, January 2025. ISSN 1996-1073. doi: 10.3390/en18133478. URL https://www.mdpi.com/1996-1073/18/13/3478.DOI
  10. [10]Delmaria Richards, Helmut Yabar, Takeshi Mizunoya, Randy Koon Koon, Gia Hong Tran, and Yannick Espere. Sustainable solar energy deployment: a multi-criteria decision-making approach for site suitability and greenhouse gas emission reduction. Environmental Science and Pollution Research, 32(4):2007–2035, January 2025. ISSN 1614-7499. doi: 10.1007/s11356-024-35669-6. URL https://doi.org/10.1007/s11356-024-35669-6.
  11. [11]Far Chen Jong and Musse Mohamud Ahmed. Multi-Criteria Decision-Making Solutions for Optimal Solar Energy Sites Identification: A Systematic Review and Analysis. IEEE Access, 12:143458–143484, 2024. ISSN 2169-3536. doi: 10.1109/ACCESS.2024.3461948. URL https://ieeexplore.ieee.org/abstract/document/10681069.
  12. [12]Alessandro Sebastianelli, Federico Serva, Andrea Ceschini, Quentin Paletta, Massimo Panella, and Bertrand Le Saux. Machine learning forecast of surface solar irradiance from meteo satellite data. Remote Sensing of Environment, 315:114431, December 2024. ISSN 0034-4257. doi: 10.1016/j.rse.2024.114431. URL https://www.sciencedirect.com/science/article/pii/S0034425724004577.DOI
  13. [13]B. Nishant Sree Reddy, Kumar Gautam, and Nikhil Pachauri. Solar potential assessment using machine learning and climate change projections for long-term energy planning. Scientific Reports, 15(1):39935, November 2025. ISSN 2045-2322. doi: 10.1038/s41598-025-23661-0. URL https://www.nature.com/articles/s41598-025-23661-0.DOI
  14. [14]Saman Nadizadeh Shorabeh, Najmeh Neysani Samany, Foad Minaei, Hamzeh Karimi Firozjaei, Mehdi Homaee, and Ali Darvishi Boloorani. A decision model based on decision tree and particle swarm optimization algorithms to identify optimal locations for solar power plants construction in Iran. Renewable Energy, 187:56–67, March 2022. ISSN 0960-1481. doi: 10.1016/j.renene.2022.01.011. URL https://www.sciencedirect.com/science/article/pii/S0960148122000118.DOI
  15. [15]Yanwei Sun, Danfeng Zhu, Ying Li, Run Wang, and Renfeng Ma. Spatial modelling the location choice of large-scale solar photovoltaic power plants: Application of interpretable machine learning techniques and the national inventory. Energy Conversion and Management, 289:117198, August 2023. ISSN 0196-8904. doi: 10.1016/j.enconman.2023.117198. URL https://www.sciencedirect.com/science/article/pii/S0196890423005447.DOI
  16. [16]Jiayue Chen, Jianxiang Shen, Shihui Zhang, Rui Wang, Zihan Zhen, Can Wang, and Wenjia Cai. Solar and wind power plant site selection in China: Machine learning-based regional and temporal probability. Applied Energy, 402:127033, January 2026. ISSN 0306-2619. doi: 10.1016/j.apenergy.2025.127033. URL https://www.sciencedirect.com/science/article/pii/S0306261925017635.DOI
  17. [17]Linwei Tao, Kiichiro Hayashi, Sangay Gyeltshen, and Yuya Shimoyama. Spatial assessment of utility-scale solar photovoltaic siting potential using machine learning approaches: A case study in Aichi prefecture, Japan. Applied Energy, 383:125329, April 2025. ISSN 0306-2619. doi: 10.1016/j.apenergy.2025.125329. URL https://www.sciencedirect.com/science/article/pii/S0306261925000595.
  18. [18]Marcel Suri, Juraj Betak, Konstantin Rosina, Daniel Chrkavy, Nada Suriova, Tomas Cebecauer, Marek Caltik, and Branislav Erdelyi. Global Photovoltaic Power Potential by Country. Technical report, Energy Sector Management Assistance Program (ESMAP), World Bank Group, Washington, D.C., 2020. URL http://documents.worldbank.org/curated/en/466331592817725242.
  19. [19]Ankita Saxena, Calum Brown, Almut Arneth, and Mark Rounsevell. Modelling the global photovoltaic potential on land and its sensitivity to climate change. Environmental Research Letters, 18(10):104017, September 2023. ISSN 1748-9326. doi: 10.1088/1748-9326/acf86f. URL https://doi.org/10.1088/1748-9326/acf86f.
  20. [20]Global Energy Monitor. Global Solar Power Tracker. 2026. URL https://globalenergymonitor.org/projects/global-solar-power-tracker. February 2026 release.
  21. [21]Rui Song, Feng Yin, Jan-Peter Muller, Adam C. Povey, Basudev Swain, Chenchen Huang, and Roy G. Grainger. Coal plants persist as a large barrier to the global solar energy transition. Nature Sustainability, 9(7):981–992, July 2026. ISSN 2398-9629. doi: 10.1038/s41893-026-01836-5. URL https://www.nature.com/articles/s41893-026-01836-5.DOI
  22. [22]Konstantin Klemmer, Esther Rolf, Marc Rußwurm, Gustau Camps-Valls, Mikolaj Czerkawski, Stefano Ermon, Alistair Francis, Nathan Jacobs, Hannah Kerner, Lester Mackey, Gengchen Mai, Oisin Mac Aodha, Markus Reichstein, Caleb Robinson, David Rolnick, Evan Shelh am er, Vincent Sitzmann, Devis Tuia, and Xiaoxiang Zhu. Earth Embeddings: Toward artificial intelligence-centric representations of our planet. IEEE Geoscience and Remote Sensing Magazine, pages 2–15, 2026. ISSN 2168-6831. doi: 10.1109/MGRS.2026.3710416. URL https://ieeexplore.ieee.org/document/11636001/.DOI
  23. [23]Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha Samsikova, Chenhui Zhang, Evan Shelahmer, Estefania Labeira, Olivia Wiles, Simon Ilyushenko, 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 and efficient global mapping from sparse label data, 2025. URL https://arxiv.org/abs/2507.02291.
  24. [24]Luc Houriez, Sebastian Pilarski, Behzad Vahedi, Ali Ahmadalipour, Teo Honda Scully, Nicholas Aflitto, David Andre, Caroline Jaffe, Martha Wedner, Rich Mazzola, Josh Jeffery, Ben Messinger, Sage McGinley-Smith, and Sarah Russell. Scalable Geospatial Data Generation Using AlphaEarth Foundations Model, 2025. URL https://arxiv.org/abs/2508.11739.
  25. [25]Xiaopeng Zeng, Weilu Sun, Mingming Jia, Zhaohui Xue, Chan Zhou, and Liqun Sun. Global high-resolution mapping of photovoltaic power plants from 2019 to 2025 using unsupervised index-based multi-source data fusion method. International Journal of Applied Earth Observation and Geoinformation, 145:105005, December 2025. ISSN 1569-8432. doi: 10.1016/j.jag.2025.105005. URL https://www.sciencedirect.com/science/article/pii/S1569843225006521.DOI
  26. [26]Rui Song and Feng Yin. Global Facility-Level Solar Photovoltaic Inventory with Energy Generation and Loss Estimates. March 2026. doi: 10.5281/zenodo.18794231. URL https://zenodo.org/records/18794231.DOI
  27. [27]Mason Phillpott, Joseph O’Connor, André Ferreira, Max Santos, Lucas Kruitwagen, and Michael Guzzardi. Solar Asset Mapper: A continuously-updated global inventory of solar energy facilities built with satellite data and machine learning. May 2024. doi: 10.5281/zenodo.11368204. URL https://zenodo.org/records/11368204.
  28. [28]Yanyao Yuan, Fenghua Liu, Guan Yang, and Miao Wang. A review of representation and similarity measurement methods for geospatial scenes. International Journal of Geographical Information Science, 40(5):1519–1546, May 2026. ISSN 1365-8816. doi: 10.1080/13658816.2025.2563697. URL https://doi.org/10.1080/13658816.2025.2563697.
  29. [29]Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The Elements of Statistical Learning, chapter 4.3, pages 106–119. Springer Series in Statistics. Springer, New York, 2nd edition, 2009.
  30. [30]Isaac Corley, Caleb Robinson, Inbal Becker-Reshef, and Juan M. Lavista Ferres. From Pixels to Patches: Pooling Strategies for Earth Embeddings, March 2026. URL http://arxiv.org/abs/2603.02800. arXiv:2603.02080 [cs.CV].
  31. [31]Steven J. Phillips, Robert P. Anderson, and Robert E. Schapire. Maximum entropy modeling of species geographic distributions. Ecological Modelling, 190(3-4):231–259, 2006. doi: 10.1016/j.ecolmodel.2005.03.026.DOI
  32. [32]Ali Rahimi and Benjamin Recht. Random Features for Large-Scale Kernel Machines. In Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007. URL https://papers.nips.cc/paper_files/paper/2007/hash/013a006f03dbc5392fefbb8f18fda755-Abstract.html.
  33. [33]Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001. doi: 10.1023/A:1010933404324.DOI
  34. [34]Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: a highly efficient gradient boosting decision tree. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS ’17, pages 3149–3157, Red Hook, NY, USA, December 2017. Curran Associates Inc. ISBN 978-1-5108-6096-4. URL https://dl.acm.org/doi/10.5555/3294996.3295074.
  35. [35]David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 323(6088):533–536, October 1986. ISSN 1476-4687. doi: 10.1038/323533a0. URL https://www.nature.com/articles/323533a0.DOI
  36. [36]Ludmila I. Kuncheva. Combining Pattern Classifiers: Fusion of Label Outputs, pages 111–149. John Wiley and Sons, Ltd., 2004. ISBN 9780471660262. doi: 10.1002/0471660264.ch4. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/0471660264.ch4.
  37. [37]Zhengpeng Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young, Madeline C. Lisaius, 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, April 2026. URL http://arxiv.org/abs/2506.20380. arXiv:2506.20380 [cs.LG].
  38. [38]Kimberley A. Corwin, Jesse Burkhardt, Chelsea A. Corr, Paul W. Stackhouse, Amit Munshi, and Emily V. Fischer. Solar energy resource availability under extreme and historical wildfire smoke conditions. Nature Communications, 16(1):245, January 2025. ISSN 2041-1723. doi: 10.1038/s41467-024-54163-8. URL https://www.nature.com/articles/s41467-024-54163-8.DOI
  39. [39]Nur Atirah Ibrahim, Sharifah Rafidah Wan Alwi, Zainuddin Abd Manan, Azizul Azri Mustafa, and Kamarizan Kidam. Climate change impact on solar system in Malaysia: Techno-economic analysis. Renewable and Sustainable Energy Reviews, 189:113901, 2024. ISSN 1364-0321. doi: 10.1016/j.rser.2023.113901. URL https://www.sciencedirect.com/science/article/pii/S1364032123007591.DOI
  40. [40]Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.arxiv.org/abs/1201.0490
  41. [41]James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(10):281–305, 2012. URL https://jmlr.org/papers/v13/bergstral2a.html.
  42. [42]Daniele Zanaga, Ruben Van De Kerchove, Dirk Daems, Wanda De Keersmaecker, Carsten Brockmann, Grit Kirches, Jan Wevers, Oliver Cartus, Maurizio Santoro, Steffen Fritz, Myroslava Lesiv, Martin Herold, Nandini-Erdene Tsendbazar, Papan Xu, Fabrizio Ramoino, and Olivier Arino. ESA WorldCover 10 m 2021 v200, 2022. URL https://doi.org/10.5281/zenodo.7254221.
  43. [43]European Space Agency and Sinergise. Copernicus Global Digital Elevation Model (GLO-30), 2021. URL https://doi.org/10.5069/G9028PQB. Distributed by OpenTopography.
  44. [44]Solargis and World Bank Group. Global Solar Atlas 2.0. https://globalsolaratlas.info, 2019. Developed and operated by Solargis s.r.o. on behalf of the World Bank Group, funded by ESMAP. Accessed 2026.
  45. [45]OpenStreetMap contributors. OpenStreetMap. https://www.openstreetmap.org, 2026. Road network from Geofabrik extracts, accessed July 2026.
  46. [46]Christopher Arderne, Conrad Zorn, Claire Nicolas, and Elco E. Koks. Predictive mapping of the global power system using open data. Scientific Data, 7(1):19, 2020. doi: 10.1038/s41597-019-0347-4.DOI
  47. [47]Center for International Earth Science Information Network (CIESIN), Columbia University. Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11, 2018. URL https://doi.org/10.7927/H49C6VHW.
  48. [48]UNEP-WCMC and IUCN. Protected Planet: The World Database on Protected Areas (WDPA). https://www.protectedplanet.net, 2026. Accessed via Google Earth Engine (WCMC/WDPA/current), 2026.

Paper details

vlnc:W_v2uss9sw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 13

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents