From Machine Learning to Large-Scale EO Products: Best Practices for Making Maps Ghjulia Sialelli1,2⋆ , Robin Young3 , Yuchang Jiang4 , Cesar Aybar5 , Linus Scheibenreif1 , Damien Robert6 , Clemens Mosig7 , Adam J. Stewart8 , Jan D. Wegner6 , Aleksis Pirinen9 , Olof Mogren9 , and Konrad Schindler1 Photogrammetry and Remote Sensing, ETH Zurich, Switzerland, 2 ETH AI Center, Zurich, Switzerland 3 Dept. of Computer Science and Technology, University of Cambridge, UK 4 Land Change Science, Swiss Federal Research Institute WSL, Switzerland 5 Asterisk Labs, UK 6 EcoVision Lab, University of Zurich, Switzerland 7 Inst. for Earth System Science and Remote Sensing, Leipzig University, Germany 8 Chair of Data Science in Earth Observation, TU Munich, Germany 9 RISE Research Institutes of Sweden · Climes, Swedish Centre for Impacts of Climate Extremes · Climate AI Nordics
arXiv:2607.24532v1 [cs.LG] 27 Jul 2026
1
Abstract. Recent years have seen a rapid expansion in the production of large-scale geospatial maps derived from Earth observation (EO) data, driven largely by advances in machine learning (ML) and large computing infrastructure. Although the barrier to generating such maps has dropped substantially, established best practices have yet to emerge, and design decisions made early in the pipeline can quietly propagate errors into the final product. Producing a technically sound and scientifically credible product remains challenging. Choices made at every stage are tightly coupled: preprocessing decisions shape the training signal, dataset design governs what the model can learn and how reliably its performance can be assessed, and global-scale inference introduces engineering challenges in compute and data access at scale, as well as artifact mitigation. Furthermore, uncertainty quantification and independent map validation each require dedicated methodological attention that is often underestimated. This paper presents a concise, end-to-end account of the recommended practices spanning the pipeline from satellite data to an operational map product. We organize the discussion around six interconnected themes: the EO data infrastructure landscape, data selection and preprocessing, ML dataset construction and model training, uncertainty quantification, map production and distribution, and validation. This paper is a condensed version of a longer guide that provides greater depth across all stages, accessible online at ghjuliasialelli.github.io/MLEO-Maps/. Keywords: Earth observation · Machine learning · Remote sensing · Best practices · Geospatial mapping ⋆
Corresponding author: [email protected]. This research was supported by the ETH AI Center through an ETH AI Center doctoral fellowship to Ghjulia Sialelli.
2
G. Sialelli et al.
1
Introduction
A growing range of global products is now routinely generated at resolutions that would have been infeasible a few years ago, enabled by advances in model capacity and the availability of large-scale compute. These products span diverse thematic areas, from human settlement [9] to vegetation structure [17, 36] and natural hazards [20]. ML has shifted the paradigm from locally calibrated models toward single, globally generalizing models trained on heterogeneous data, attracting a growing number of practitioners from both the EO and ML communities. Yet, the resulting workflow introduces challenges that are largely new to the ML community. When considering the full workflow for a global map, every stage presents domain-specific challenges that are tightly coupled: preprocessing decisions shape the training signal, dataset design governs what the model can learn and how reliably its performance can be assessed, and global-scale inference introduces engineering challenges in compute, data movement, and artifact mitigation. Furthermore, uncertainty quantification and independent map validation each require dedicated methodological attention that is often underestimated. Much of this practical knowledge remains dispersed, siloed within specific application domains, making it difficult for practitioners to learn from one another across fields. This paper is the result of a broad effort to engage with practitioners across the EO and ML communities. We distill common practices, recurring pitfalls, and recommended approaches into a concise account organized around six interconnected themes: the EO data infrastructure landscape, data selection and preprocessing, ML dataset construction and model training, uncertainty quantification, map production and distribution, and validation. We hope this document serves as both a practical reference and an invitation for continued exchange, and refer readers to the online version for additional depth.
2
EO Data Infrastructure Landscape
Researchers must navigate a complex ecosystem of satellite data sources, derived data products, and access platforms. The choices made at the data access level have downstream consequences that are easy to underestimate: different providers and platforms expose different processing levels and retrieval capabilities, which can quietly constrain the rest of the pipeline. Curated ML-ready EO datasets [8, 35] offer an accessible entry point, but adopting one means inheriting its creators’ design choices (spatiotemporal coverage, sampling strategy, processing pipeline) which may not match the requirements of the downstream task. Hence, most practitioners create their own EO dataset. When starting from satellite observations, a unifying constraint is data gravity: modern EO archives have grown to a scale where even freely available data is often impractical to move, and the position of compute relative to data has become a primary design factor. This is why many large-scale workflows rely
Best Practices for EO Maps
3
on cloud platforms that co-locate data and compute on the cloud (more on that in §6). The access landscape ranges from canonical providers (ESA CDSE, USGS EarthExplorer, NASA Earthdata) to re-distribution platforms (Google Earth Engine, Microsoft Planetary Computer, AWS S3), each with distinct trade-offs in server-side operations, suitability for large-scale downloads, and collocated compute availability (see Tab. 1 in the appendix). A practical pattern is to treat access as a two-stage process: model development typically involves modest data volumes downloadable from almost any provider, while inference at scale is dominated by data movement. However, processing pipelines may differ between platforms. For example, Google Earth Engine automatically corrects for the radiometric offset that was introduced for Sentinel-2 data post Processing Baseline 04.00; they also apply a full preprocessing chain (thermal noise removal, data calibration, multi-looking and range-doppler terrain correction) to Sentinel1 GRD data, rather than redistributing the GRD product as given by ESA. This can lead to distribution shifts when a model trained on data from one provider is deployed on data from another.
3
Data Selection and Preprocessing
Data selection and preprocessing choices are not neutral; they encode assumptions that propagate into the training pipeline and ultimately shape the final map product. While data selection is primarily guided by the target domain and task, it is also shaped by practical considerations such as cloud contamination, misalignment, and sensor artifacts. These issues can either be handled explicitly before the data reaches the model, or left for the model to learn to handle through exposure to natural variation and targeted data augmentation (see §4), saving preprocessing cost but consuming model capacity on problems that are not the target task. Beyond individual observations, how training samples are distributed geographically and temporally determines what the model can learn and how well it generalizes. Furthermore, whatever pipeline is adopted for training must also be reproduced at inference time over the full mapped domain, so every additional processing step and temporal input compounds the operational cost of deployment. While aggressive filtering is acceptable when curating a training dataset, inference must still produce predictions over the entire deployment area, even in persistently difficult conditions. Data selection. Individual observations vary in usability. For optical imagery, a common filter consists in discarding products above a certain threshold of cloud cover. However, for persistently cloudy regions, this risks discarding valuable non-cloudy pixels. Sentinel-2 products, specifically, exhibit artifacts from incomplete orbits and split partial products (see Appendix B). Landsat 7 scenes acquired after the 2003 scan-line corrector failure exhibit systematic data gaps. If not handled explicitly, these can cause pixel duplication, bias dataset statistics, and produce visible seams or gaps in the final map; issues that are often only discovered at deployment time, forcing costly reprocessing. SAR imagery
4
G. Sialelli et al.
is not affected by clouds but its side-looking geometry introduces geometric distortions (foreshortening, layover, shadow) [7]. Ascending and descending passes image the same terrain from opposite look angles, so combining both can fill distortion gaps in mountainous regions. Furthermore, track selection can be guided by sub-swath incidence angles and DEM-derived slope alignment. Spatio-temporal strategy. How training data is sampled geographically has a direct effect on what the model learns and how well it generalizes. Reference data coverage is typically uneven worldwide, e.g., OpenStreetMap offers more coverage in the Global North than the Global South [11], reflecting not only disparities in accessibility, but also the priorities of the communities that create and curate tehese datasets. Models trained without accounting for this imbalance risk learning a skewed representation of the Earth’s surface. On the temporal side, when producing a map over a given period, a few broad strategies emerge: single time-steps (most cost-efficient but sacrifice temporal information), time-series (richest, capturing seasonal variation, but memory-intensive [28]), composites (pragmatic middle ground, distilling a period into a single time-step [27]), and time series of composites [9,12,23]. The choice directly affects the volume of data that must be downloaded or streamed at inference time, which at global scale can be the binding constraint. Preprocessing. Each correction applied during training must be replicated identically at inference time, so the complexity of the preprocessing chain directly translates into operational cost at scale. Typical steps vary by sensor and, crucially, between providers. Sentinel-1 GRD products, for instance, may require thermal noise removal, radiometric calibration, speckle reduction, and geometric/radiometric terrain correction; while Sentinel-2 L2A products require radiometric offset handling (see Appendix C) and optional BRDF normalization [21].
4
Machine Learning Dataset Construction and Model Training
With preprocessed data in hand, the focus shifts to the ML stages of the pipeline. Constructing a training dataset from EO imagery introduces challenges that have no direct analogue in standard computer vision, from spatial autocorrelation that inflates evaluation metrics, to gridding choices that silently bias sampling. On the modeling side, architecture and training decisions must be made with globalscale inference in mind, since a model that cannot run efficiently over the full mapped domain is of limited operational value. Data formats. The data format used to store the dataset is a creation-time decision that propagates through the entire pipeline, defining how array payloads are physically laid out, how metadata are encoded, and how chunks are addressed. Two broad storage patterns exist: container formats with an internal index (HDF5, NetCDF4, GeoTIFF, COG) and key-value chunk stores (Zarr). We provide a detailed comparison in Table 2 (Appendix D).
Best Practices for EO Maps
5
Dataset splits. Spatial autocorrelation in labels inflates apparent performance if not handled at the splitting stage [29, 30]. We illustrate some approaches in Fig. 2 (Appendix E). Spatial K-fold block cross-validation divides the area into non-overlapping spatial blocks and assigns each to a fold. To avoid cross-fold contamination, spatial buffers between blocks can be enforced, as in [32], if the blocks cannot be large enough. A pragmatic approach consists in enforcing a distance between folds that exceeds the spatial autocorrelation range of the target variable, as in [19, 26]. Spatial gridding. A grid defines the spatial index over which data is sampled and split. Candidate grids, ranging from Web Mercator and UTM/MGRS to Discrete Global Grid Systems such as HEALPix, differ in whether they are equal-area (every cell represents the same ground area), conformal (preserves local angles and shapes such that a square pixel corresponds to a square patch on the ground), and globally continuous (every point on the sphere belongs to exactly one cell). These properties matter for unbiased sampling, accurate area estimation, and dataset splitting. We provide an overview in Table 3 (Appendix G). We advocate decoupling the indexing strategy (how to select samples) from the patch projection (the coordinate system each sample lives in). Indexing should use a global grid that minimizes over-/under-sampling and overlap; the patch projection only needs to minimize distortion at the scale of a single patch. Model design and training. Deploying a model over the full mapped domain means that every forward pass performed during training will be repeated billions of times at inference. This makes architecture choice an operational decision as much as a modelling one: Vision Transformers and foundation models are increasingly explored for EO tasks [6,14], but convolutional encoder-decoders remain the workhorse of operational pipelines [9,18] precisely because their higher throughput makes wall-to-wall inference tractable. Furthermore, the training procedure interacts with the preprocessing choices discussed in §3 via data augmentation. Rather than explicitly correcting for sensor artifacts in preprocessing, or missing modalities, one can instead simulate these corruptions during training so the model learns to be invariant to them. To enhance spectral robustness, noise can be added to Sentinel-2 surface reflectance values [2,22]. Temporal augmentations, such as dropping timesteps, can simulate cloud-induced data gaps or missing modalities. Simulating narrow no-data seams in Sentinel-1 GRD mosaics [3], as well as modeling swath boundaries, misregistration, and detector failures [25] can build resilience to specific sensor artifacts. Model validation. Held-out test metrics (computed on the spatial splits described above) measure how well the model generalizes to new locations drawn from the same data collection process. This is essential for architecture selection, hyperparameter tuning, and diagnosing failure modes, but it does not answer whether the resulting map is accurate across its full extent, especially with opportunistic data collection. We therefore distinguish model validation (estimating predictive performance on data drawn from the same collection as the training set)
6
G. Sialelli et al.
from map validation (assessing the final product against independently collected reference data with a known sampling design, see §7).
5
Uncertainty Quantification
International frameworks for climate reporting require quantified uncertainty [5], and carbon credit markets depend on it [10]. Uncertainty is what transforms a map from a static product into a scientifically credible tool that guides action. Sources. Total uncertainty arises from upstream (input) uncertainty (sensor noise, atmospheric correction errors, cloud mask failures), label uncertainty (measurement error in reference data), and model (epistemic) uncertainty. Many uncertainty quantification (UQ) methods address only model uncertainty; the gap should be made explicit. Methods. No method is simultaneously cheap, scalable, and well-calibrated. Gaussian Processes are calibrated (if the kernel is correct) but do not scale. Deep ensembles [16] scale but are expensive and can be miscalibrated for spatial data: members trained on the same spatially autocorrelated data agree where they are collectively wrong. Heteroskedastic regression is cheap but blind to epistemic uncertainty. Conformal prediction [38] provides distribution-free coverage guarantees as a cheap post-hoc wrapper around any predictor, but the guarantee is marginal. For classification tasks, conformalized quantile regression [31] improves on this by calibrating model-produced prediction intervals, but coverage holds on average over the test distribution, not for any specific subset of it. Users should articulate the intended use of uncertainty before choosing a method. Calibration. Reporting uncertainty without verifying calibration is akin to reporting accuracy without a test set. We recommend coverage and reliability plots, standardized residuals (z-scores with ideal mean 0 and std 1), and proper scoring rules (interval score, CRPS). Spatial structure. Map errors are spatially correlated. √ If pixel errors were independent, aggregate uncertainty would shrink as 1/ n with the number of pixels. But when errors are positively correlated, they do not cancel upon averaging. Wadoux et al . [39] demonstrated that several high-profile studies reported aggregate uncertainty without accounting for spatial autocorrelation, leading to substantially underestimated confidence intervals. Johnson et al . [15] found that residual variance from spatial correlation dominated all other sources of uncertainty for biomass aggregated to ownership parcels. This remains one of the largest gaps between current practice and what is needed for rigorous area-based reporting. As a practical minimum, we recommend characterizing residual spatial correlation in a small number of ecologically distinct validation regions.
Best Practices for EO Maps
6
7
Map Production and Distribution
Producing a map from a trained model is often treated as an engineering detail, but the choices made at this stage (how to orchestrate inference, how to handle tiling artifacts, how to compress and distribute the result) directly affect the quality and usability of the final product. Compute environments. The computational strategy for global-scale inference is shaped by the data gravity constraint introduced in §A: at petabyte scale, where the data is stored relative to compute determines the feasibility of the pipeline. We identify two regimes. In the data-to-compute regime, data is downloaded to (or streamed on) institutional clusters. This regime is attractive when compute is subsidized, but bounded by network bandwidth, provider-side rate limits, and egress costs. In the compute-to-data regime, compute is provisioned in the same cloud region where the data resides, eliminating transfer penalties. The cost of cloud compute can however escalate quickly. Spot instances (while substantially cheaper) increase complexity as they require checkpoint-and-resume pipelines to handle preemption. Artifact mitigation. Due to memory constraints, EO scenes are partitioned into patches for inference that require merging in postprocessing. Further, models that work with spatial context may exhibit spatial bias: predictions degrade toward patch borders due to zero-padding, non-unary strides, and windowed attention [41]. Two main strategies [13] address this: (a) overlapping patches with weighted blending, where redundant outputs are merged using distance-based weights [23, 26]; and (b) padding and cropping, where each patch is expanded with context and only center predictions are retained [3, 27]. We illustrate these approaches in Fig. 3 (Appendix F). Sharing. To maximize impact, map products should adhere to the FAIR principles [40] with rich metadata and persistent identifiers, and should support both large-scale batch downloads and interactive visualization. Cloud Optimized GeoTIFFs and Zarr are both cloud-native formats seeing wide adoption, though each has limitations (Table 2). Indexing with SpatioTemporal Asset Catalog (STAC, a standardized specification for describing geospatial assets with consistent spatial and temporal metadata) makes products discoverable and searchable across archives.
7
Map Validation
Map validation assesses the final product against systematically and independently collected reference data covering the full mapped domain, whereas model validation (§4) is limited to estimating predictive performance on data drawn from the same collection as the training set. Such a systematic evaluation is what stands between map production and map credibility, yet accuracy assessment methodology is frequently under-reported [34].
8
G. Sialelli et al.
Common sources of confusion. A held-out test set drawn from the same data collection process as the training set shares its biases and coverage gaps; it measures generalization within that distribution, not quality of the map as a whole. Furthermore, uncertainty quantification methods have limitations (see §5), and comparison against another map only measures inter-product consistency. None of these approaches are substitutes for independent validation, which requires reference data that is independent of the training pipeline, collected through a separate process, ideally with a sampling design that covers the full mapped domain. Probability sampling and design-based assessment. The gold standard for map accuracy assessment is a probability-based sampling design where every location in the mapped domain has a known, non-zero chance of being selected for verification [33]. This allows unbiased estimation of accuracy over the full map, not just at convenient or available locations. In practice, this means drawing a sample (e.g., stratified random), collecting reference labels for those locations, and constructing area-weighted confusion matrices that account for differences in how densely each stratum was sampled [24]. Per-class accuracy should be reported from both the user’s perspective (how often a predicted class is correct, or precision) and the producer’s perspective (how often an actual class is detected, or recall) and accompanied by confidence intervals (see §5). Beyond probability sampling. For some products, e.g., continuous-values maps like biomass, strict probability sampling may not be practical. Alternative validation using opportunistically sampled observations can provide valuable verification, but practitioners should identify which ecoregions, land cover types, and target variable ranges are most underrepresented in the reference data, and prioritize additional data collection accordingly [4]. The limitations should also be stated explicitly: such methods measure agreement at sampled locations rather than accuracy over the full map extent [37].
8
Conclusion
This paper has traced the entire pipeline for producing large-scale maps from EO data, exposing the interdependencies between stages that are easy to overlook when each is treated in isolation. The practical knowledge needed to navigate these stages is often dispersed across communities; our aim has been to bring it together in one place. We have highlighted that the choice of data provider is not neutral, that data selection and preprocessing decisions encode assumptions the model will inherit, and that whatever chain is adopted must be sustainable at inference scale. On the ML side, spatial autocorrelation demands blocked splits to avoid inflated metrics, gridding choices affect both sampling and evaluation validity, and data augmentation can absorb some of the burden that would otherwise fall on preprocessing. Uncertainty quantification is not optional, but no existing method is simultaneously cheap, scalable, and well-calibrated; and the
Best Practices for EO Maps
9
spatial correlation of map errors means that naive aggregation dramatically understates uncertainty at regional scales. Finally, model evaluation is not map validation: accuracy assessment (through probability sampling for classification tasks, or carefully characterized alternative approaches for continuous-valued products) remains essential, and its limitations should be stated rather than ignored. For a complete treatment of each stage, we refer the reader to the online guide at ghjuliasialelli.github.io/ML-EO-Maps/.
Acknowledgements This research was supported by the ETH AI Center through an ETH AI Center doctoral fellowship to Ghjulia Sialelli. We thank Isabelle Tingzon, Chiara Ceccobello, and Sebastian Hafner (RISE Research Institutes of Sweden), Johannes Reiche, Robert N. Masolele, and Nandika Tsendbazar (Wageningen University), Nico Lang (University of Copenhagen), Jan Pauls (University of Münster), Zhengpeng Feng (University of Cambridge), Daniel Lusk (University of Freiburg), Martin Schwartz (LSCE), Christelle Vancutsem (EU Joint Research Centre), Peter Potapov (World Resources Institute), Kristof Van Tricht (VITO), Maurizio Santoro (Gamma Remote Sensing), Marcin Kluczek (CloudFerro), Louis de Vitry (Kanop), Mikołaj Czerkawski (Asterisk Labs), Valerie J. Pasquarella (Google), Henry Herzog (Allen Institute for AI), Caleb Robinson (Microsoft AI for Good Research Lab), Gabriel Belouze, Isaac Corley, and Harald Kristen for generously sharing their experience and insights during the preparation of this work. Their input helped shape several of the practical recommendations presented here, though responsibility for any errors or omissions remains with the authors.
References 1. Bauer-Marschallinger, B., Falkner, K.: Wasting petabytes: A survey of the Sentinel2 UTM tiling grid and its spatial overhead. ISPRS Journal of Photogrammetry and Remote Sensing 202, 682–690 (Aug 2023). https://doi.org/10.1016/j. isprsjprs.2023.07.015, http://dx.doi.org/10.1016/j.isprsjprs.2023.07. 015 2. Brown, C.F., Brumby, S.P., Guzder-Williams, B., Birch, T., Hyde, S.B., Mazzariello, J., Czerwinski, W., Pasquarella, V.J., Haertel, R., Ilyushchenko, S., Schwehr, K., Weisse, M., Stolle, F., Hanson, C., Guinan, O., Moore, R., Tait, A.M.: Dynamic World, Near real-time global 10m land use land cover mapping. Scientific Data 9(1) (Jun 2022). https://doi.org/10.1038/s41597-022-01307-4, http://dx.doi.org/10.1038/s41597-022-01307-4 3. Brown, C.F., Kazmierski, M.R., Pasquarella, V.J., Rucklidge, W.J., Samsikova, M., Zhang, C., Shelhamer, E., Lahera, E., Wiles, O., Ilyushchenko, S., Gorelick, N., Zhang, L.L., Alj, S., Schechter, E., Askay, S., Guinan, O., Moore, R., Boukouvalas, A., Kohli, P.: AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv preprint arXiv:2507.22291 (2025), https://arxiv.org/abs/2507.22291
10
G. Sialelli et al.
4. Duncanson, L., Armston, J., Disney, M., Avitabile, V., Barbier, N., Calders, K., Carter, S., Chave, J., Herold, M., MacBean, N., McRoberts, R., Minor, D., Paul, K., Réjou-Méchain, M., Roxburgh, S., Williams, M., Albinet, C., Baker, T., Bartholomeus, H., Bastin, J.F., Coomes, D., Crowther, T., Davies, S., de Bruin, S., De Kauwe, M., Domke, G., Falkowski, M., Fatoyinbo, L., Goetz, S., Jantz, P., Jonckheere, I., Jucker, T., Kay, H., Kellner, J., Labriere, N., Lucas, R., Morsdorf, F., Phillips, O.L., Quegan, S., Saatchi, S., Schaaf, C., Schepaschenko, D., Scipal, K., Stovall, A., Thiel, C., Wulder, M.A., Camacho, F., Nickeson, J., Roman, M., Margolis, H.: Global Aboveground Biomass Product Validation Best Practices Protocol. In: Duncanson, L., Disney, M., Armston, J., Minor, D., Camacho, F., Nickeson, J. (eds.) Best Practice Protocol for Satellite Derived Land Product Validation, p. 222. Land Product Validation Subgroup (WGCV/CEOS) (2020). https://doi.org/10.5067/doc/ceoswgcv/lpv/agb.001, version 1.0 5. Eggleston, H.S., Buendia, L., Miwa, K., Ngara, T., Tanabe, K. (eds.): 2006 IPCC Guidelines for National Greenhouse Gas Inventories. Institute for Global Environmental Strategies (IGES) for the IPCC, Hayama, Japan (2006), https: //www.ipcc-nggip.iges.or.jp/public/2006gl/ 6. Feng, Z., Atzberger, C., Jaffer, S., Knezevic, J., Sormunen, S., Young, R., Lisaius, M.C., Immitzer, M., Jackson, T., Ball, J., Coomes, D.A., Madhavapeddy, A., Blake, A., Keshav, S.: TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis. arXiv preprint arXiv:2506.20380 (2025), https:// arxiv.org/abs/2506.20380 7. Flores-Anderson, A., Herndon, K., Thapa, R., Cherrington, E.: The SAR Handbook: Comprehensive Methodologies for Forest Monitoring and Biomass Estimation (04 2019). https://doi.org/10.25966/nr2c-s697 8. Francis, A., Czerkawski, M.: Major TOM: Expandable Datasets for Earth Observation. In: IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium. pp. 2935–2940 (2024). https://doi.org/10.1109/IGARSS53475. 2024.10640760 9. Glazer, T., Hacheme, G.Q., Zaytar, A., Marotti, L., Michaels, A., Tadesse, G.A., White, K., Dodhia, R., Zolli, A., Becker-Reshef, I., Lavista Ferres, J.M., Robinson, C.: TEMPO: Global Temporal Building Density and Height Estimation from Satellite Imagery. arXiv preprint arXiv:2511.12104 (2025), https://arxiv.org/ abs/2511.12104 10. Haya, B., Cullenward, D., Strong, A.L., Grubert, E., Heilmayr, R., Sivas, D.A., Wara, M.: Managing uncertainty in carbon offsets: insights from California’s standardized approach. Climate Policy 20(9), 1112–1126 (2020). https://doi.org/ 10.1080/14693062.2020.1781035 11. Herfort, B., Lautenbach, S., Porto de Albuquerque, J., Anderson, J., Zipf, A.: A spatio-temporal analysis investigating completeness and inequalities of global urban building data in openstreetmap. Nature Communications 14(1) (Jul 2023). https://doi.org/10.1038/s41467-023-39698-6, http://dx.doi.org/10.1038/ s41467-023-39698-6 12. Herzog, H., Bastani, F., Zhang, Y., Tseng, G., Redmon, J., Sablon, H., Park, R., Morrison, J., Buraczynski, A., Farley, K., Hansen, J., Howe, A., Johnson, P.A., Otterlee, M., Schmitt, T., Pitelka, H., Daspit, S., Ratner, R., Wilhelm, C., Wood, S., Jacobi, M., Kerner, H., Shelhamer, E., Farhadi, A., Krishna, R., Beukema, P.: OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation. arXiv preprint arXiv:2511.13655 (2025), https://arxiv.org/abs/2511.13655
Best Practices for EO Maps
11
13. Huang, B., Reichman, D., Collins, L.M., Bradbury, K., Malof, J.M.: Tiling and Stitching Segmentation Output for Remote Sensing: Basic Challenges and Recommendations. arXiv preprint arXiv:1805.12219 (2018), https://arxiv.org/abs/ 1805.12219 14. Jakubik, J., Roy, S., Phillips, C.E., Fraccaro, P., Godwin, D., Zadrozny, B., Szwarcman, D., Gomes, C., Nyirjesy, G., Edwards, B., Kimura, D., Simumba, N., Chu, L., Mukkavilli, S.K., Lambhate, D., Das, K., Bangalore, R., Oliveira, D., Muszynski, M., Ankur, K., Ramasubramanian, M., Gurung, I., Khallaghi, S., Li, H., Cecil, M., Ahmadi, M., Kordi, F., Alemohammad, H., Maskey, M., Ganti, R., Weldemariam, K., Ramachandran, R.: Foundation Models for Generalist Geospatial Artificial Intelligence (2023). https://doi.org/10.48550/ARXIV.2310.18660, https://arxiv.org/abs/2310.18660 15. Johnson, L.K., Domke, G.M., Stehman, S.V., Mahoney, M.J., Beier, C.M.: From pixels to parcels: Flexible, practical small-area uncertainty estimation for spatial averages obtained from aboveground biomass maps. Remote Sensing of Environment 330, 114951 (2025). https://doi.org/10.1016/j.rse.2025.114951, https://doi.org/10.1016/j.rse.2025.114951 16. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. In: Advances in Neural Information Processing Systems. pp. 6405–6416 (2017), https://proceedings.neurips.cc/ paper _ files / paper / 2017 / file / 9ef2ed4b7fd2c810847ffa5fa85bce38 - Paper . pdf 17. Lang, N., Jetz, W., Schindler, K., Wegner, J.D.: A High-Resolution Canopy Height Model of the Earth. Nature Ecology & Evolution 7, 1778–1789 (2023). https: //doi.org/10.1038/s41559-023-02206-6, https://doi.org/10.1038/s41559023-02206-6 18. Liu, S., Brandt, M., Nord-Larsen, T., Chave, J., Reiner, F., Lang, N., Tong, X., Ciais, P., Igel, C., Pascual, A., Guerra-Hernandez, J., Li, S., Mugabowindekwe, M., Saatchi, S., Yue, Y., Chen, Z., Fensholt, R.: The overlooked contribution of trees outside forests to tree cover and woody biomass across Europe. Science Advances 9(37) (Sep 2023). https://doi.org/10.1126/sciadv.adh4097, http://dx.doi. org/10.1126/sciadv.adh4097 19. Lusk, D., Wolf, S., Svidzinska, D., Dormann, C.F., Kattge, J., Bruelheide, H., Sabatini, F.M., Damasceno, G., Moreno Martínez, Á., Violle, C., Hending, D., Hähn, G.J.A., Tabeni, S., Phartyal, S., Gonçalves, F., Kreft, H., Schmidt, M., Chen, H., Güler, B., Dolezal, J., Pielech, R., Guido, A., Dwyer, C., Napoleone, F., Willie, J., Gasper, A.L., Macía, M.J., Chytry, M., Lenoir, J., Thakur, D., Dengler, J., Świerszcz, S., Altman, J., Mucina, L., Nerlekar, A.N., Kakinuma, K., Rawat, P., Stančić, Z., Testolin, R., Hatim, M.Z., Rodrigues, F., Homeier, J., Marques, M.C.M., McCarthy, J.K., El-Sheikh, M.A., Korznikov, K., Gerberding, K., Kattenborn, T.: Crowdsourced biodiversity monitoring fills gaps in global plant trait mapping. Nature Communications 17(1) (Jan 2026). https://doi.org/10.1038/s41467-02668996-y, http://dx.doi.org/10.1038/s41467-026-68996-y 20. Misra, A., White, K., Nsutezo, S.F., Straka, W., Lavista, J.: Mapping global floods with 10 years of satellite radar data. Nature Communications 16(1) (Jul 2025). https://doi.org/10.1038/s41467-025-60973-1, http://dx.doi.org/10.1038/ s41467-025-60973-1 21. Montero, D., Mahecha, M.D., Aybar, C., Mosig, C., Wieneke, S.: Facilitating advanced Sentinel-2 analysis through a simplified computation of Nadir BRDF Adjusted Reflectance. The International Archives of the Photogrammetry, Remote
12
G. Sialelli et al.
Sensing and Spatial Information Sciences XLVIII-4/W12-2024, 105–112 (Jun 2024). https://doi.org/10.5194/isprs-archives-xlviii-4-w12-2024-1052024, http://dx.doi.org/10.5194/isprs-archives-XLVIII-4-W12-2024-1052024 22. Mosig, C., Kattenborn, T., Montero Loaiza, D., Vanja-Jehle, J., Brandt, J., Jacobs, N., Khanal, S., Xing, E., Schwartz, M., Muller-Landau, H.C., Beloiu, M., Bozzini, A., Cheng, Y., Ganz, K., Grüning, B., Hartmann, H., Hempel, J., Horion, S., Junttila, S., Korznikov, K., Kraemer, G., Mönks, M., Nardi, D., Neumeier, P., Schmid, J., Soltani, S., Therese-Schmehl, M., Veitch-Michaelis, J., Mahecha, M.: Sub-pixel mapping of disturbance and tree mortality dynamics from Sentinel-2 time series around the globe (Feb 2026). https://doi.org/10.31223/x5b18w, http://dx.doi.org/10.31223/X5B18W 23. Neumann, M., Raichuk, A., Jiang, Y., Rey, M., Stanimirova, R., Sims, M.J., Carter, S., Goldman, E., Anderson, K., Poklukar, P., Tarrio, K., Lesiv, M., Fritz, S., Clinton, N., Stanton, C., Morris, D., Purves, D.: Natural forests of the world – a 2020 baseline for deforestation and degradation monitoring. Scientific Data 12(1) (Nov 2025). https://doi.org/10.1038/s41597-025-06097-z, http://dx.doi.org/10.1038/s41597-025-06097-z 24. Olofsson, P., Foody, G.M., Herold, M., Stehman, S.V., Woodcock, C.E., Wulder, M.A.: Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment 148, 42–57 (May 2014). https://doi.org/10. 1016/j.rse.2014.02.015, http://dx.doi.org/10.1016/j.rse.2014.02.015 25. Pasquarella, V.J., Brown, C.F., Czerwinski, W., Rucklidge, W.J.: Comprehensive quality assessment of optical satellite imagery using weakly supervised video learning. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). p. 2125–2135. IEEE (Jun 2023). https://doi.org/10. 1109/cvprw59228.2023.00206, http://dx.doi.org/10.1109/CVPRW59228.2023. 00206 26. Pauls, J., Schrödter, K., Ligensa, S., Schwartz, M., Turan, B., Zimmer, M., Saatchi, S., Pokutta, S., Ciais, P., Gieseke, F.: ECHOSAT: Estimating Canopy Height Over Space and Time. arXiv preprint arXiv:2602.21421 (2026), https://arxiv.org/ abs/2602.21421 27. Pauls, J., Zimmer, M., Kelly, U.M., Schwartz, M., Saatchi, S., Ciais, P., Pokutta, S., Brandt, M., Gieseke, F.: Estimating Canopy Height at Scale. arXiv preprint arXiv:2406.01076 (2024), https://arxiv.org/abs/2406.01076 28. Pauls, J., Zimmer, M., Turan, B., Saatchi, S., Ciais, P., Pokutta, S., Gieseke, F.: Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation. arXiv preprint arXiv:2501.19328 (2025), https://arxiv.org/abs/2501.19328 29. Ploton, P., Mortier, F., Réjou-Méchain, M., Barbier, N., Picard, N., Rossi, V., Dormann, C., Cornu, G., Viennois, G., Bayol, N., Lyapustin, A., Gourlet-Fleury, S., Pélissier, R.: Spatial validation reveals poor predictive performance of largescale ecological mapping models. Nat. Commun. 11(1), 4540 (Sep 2020) 30. Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita, G., Hauenstein, S., Lahoz-Monfort, J.J., Schröder, B., Thuiller, W., Warton, D.I., Wintle, B.A., Hartig, F., Dormann, C.F.: Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8), 913–929 (Mar 2017). https://doi.org/10.1111/ecog.02881, http://dx.doi.org/10. 1111/ecog.02881
Best Practices for EO Maps
13
31. Romano, Y., Patterson, E., Candès, E.J.: Conformalized Quantile Regression. In: Advances in Neural Information Processing Systems. pp. 3543–3553 (2019), https: //papers.neurips.cc/paper/8613-conformalized-quantile-regression.pdf 32. Sainte Fare Garnot, V., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. ICCV (2021) 33. Stehman, S.V., Czaplewski, R.L.: Design and analysis for thematic map accuracy assessment: fundamental principles. Remote sensing of environment 64(3), 331–344 (1998) 34. Stehman, S.V., Foody, G.M.: Key issues in rigorous accuracy assessment of land cover products. Remote Sensing of Environment 231, 111199 (Sep 2019). https: //doi.org/10.1016/j.rse.2019.05.018, http://dx.doi.org/10.1016/j.rse. 2019.05.018 35. Stewart, A.J., Robinson, C., Corley, I.A., Ortiz, A., Lavista Ferres, J.M., Banerjee, A.: TorchGeo: Deep Learning With Geospatial Data. ACM Transactions on Spatial Algorithms and Systems 11(4), 1–28 (Aug 2025) 36. Tolan, J., Yang, H.I., Nosarzewski, B., Couairon, G., Vo, H.V., Brandt, J., Spore, J., Majumdar, S., Haziza, D., Vamaraju, J., Moutakanni, T., Bojanowski, P., Johns, T., White, B., Tiecke, T., Couprie, C.: Very high resolution canopy height maps from rgb imagery using self-supervised vision transformer and convolutional decoder trained on aerial lidar. Remote Sensing of Environment 300, 113888 (Jan 2024). https://doi.org/10.1016/j.rse.2023.113888, http://dx.doi.org/10. 1016/j.rse.2023.113888 37. Tyukavina, A., Stehman, S.V., Pickens, A.H., Potapov, P., Hansen, M.C.: Practical global sampling methods for estimating area and map accuracy of land cover and change. Remote Sensing of Environment 324, 114714 (2025) 38. Vovk, V., Gammerman, A., Shafer, G.: Algorithmic Learning in a Random World. Springer New York, NY, 1 edn. (2005). https://doi.org/10.1007/b106715 39. Wadoux, A.C., Heuvelink, G.B.M.: Uncertainty of Spatial Averages and Totals of Natural Resource Maps. Methods in Ecology and Evolution 14, 1320–1332 (2023). https://doi.org/10.1111/2041-210X.14106, https://doi.org/10.1111/2041210X.14106 40. Wilkinson, M.D., Dumontier, M., Aalbersberg, I.J., Appleton, G., Axton, M., Baak, A., Blomberg, N., Boiten, J.W., da Silva Santos, L.B., Bourne, P.E., Bouwman, J., Brookes, A.J., Clark, T., Crosas, M., Dillo, I., Dumon, O., Edmunds, S., Evelo, C.T., Finkers, R., Gonzalez-Beltran, A., Gray, A.J., Groth, P., Goble, C., Grethe, J.S., Heringa, J., ’t Hoen, P.A., Hooft, R., Kuhn, T., Kok, R., Kok, J., Lusher, S.J., Martone, M.E., Mons, A., Packer, A.L., Persson, B., Rocca-Serra, P., Roos, M., van Schaik, R., Sansone, S.A., Schultes, E., Sengstag, T., Slater, T., Strawn, G., Swertz, M.A., Thompson, M., van der Lei, J., van Mulligen, E., Velterop, J., Waagmeester, A., Wittenburg, P., Wolstencroft, K., Zhao, J., Mons, B.: The fair guiding principles for scientific data management and stewardship. Scientific Data 3(1) (Mar 2016). https://doi.org/10.1038/sdata.2016.18, http://dx.doi.org/10.1038/sdata.2016.18 41. Zhang, R.: Making Convolutional Networks Shift-Invariant Again (2019). https: //doi.org/10.48550/ARXIV.1904.11486, https://arxiv.org/abs/1904.11486
14
A
G. Sialelli et al.
EO Data Providers
Table 1: Overview of major Earth observation data providers. Online ops: whether server-side processing is exposed through an API. Bulk egress: feasibility of largescale archive download (✓ = supported, ≈ = requires special arrangements, × = not suited). Collocated compute: whether large GPU/TPU pools are available in the same cloud region (✓ = hyperscaler-class, ≈ = limited). *GEE data is not directly user-accessible at the storage layer. Provider
Host
Microsoft Planetary Azure (Netherlands) Computer CDSE CloudFerro (PL/DE) NASA Earthdata AWS us-west-2 Source Cooperative AWS us-west-2 AWS Open Data AWS (varies) Google Earth En- GCS* gine SentinelHub AWS eu-central-1 USGS AWS us-west-2
B
Online ops Bulk egress Collocated compute ×
✓
✓
✓
✓
≈
× × × ✓
✓ ✓ ✓ ×
✓ ✓ ✓ ≈
✓ ×
× ✓
≈ ✓
Sentinel-2 Artifacts
Because the Sentinel-2 distribution unit (granule) is not geometrically aligned with the sampling unit (orbit), each tile is covered by multiple orbits, some of which encompass it entirely. In such cases, the appropriate orbit can be identified from the NODATA_PIXEL_PERCENTAGE in product metadata: it should be 0.0 for the orbit that fully covers the tile. Most tiles, however, are not fully covered by any single orbit and require mosaicking data from multiple orbits acquired at different times. A more unusual case occurs when a tile that should be fully covered by a single orbit yields two products sharing the same date, orbit, and footprint. When processing baselines are identical, this indicates a data strip split: a single continuous observation broken into segments during ground processing. These must be mosaicked to reconstruct the full area. Addressing these issues is important to avoid pixel duplication, which biases any statistics computed over the data—whether during compositing or when calculating dataset statistics [1]. Fig. 1 illustrates these three cases.
C
Sentinel-2 Radiometric Processing
Since the deployment of Sentinel-2 Processing Baseline (PB) 04.00 in January 2022, a radiometric offset is added to the digital number values of all bands.
Best Practices for EO Maps
15
Fig. 1: Sentinel-2 artifacts. Top: L2A products for tile 31TGN from two orbits on the same date—orbit 008 (left) partially covers the tile while orbit 108 (right) covers it fully. Middle: Products for tile 30SUH from two sensors on consecutive days, each covering a complementary portion. Bottom: Two L2A products for tile 30STH from the same orbit and date, resulting from a data strip split. Images acquired using the Copernicus Browser.
For Sentinel-2 L1C data, the conversion to reflectance is given by Equation 1, where ρT OA is the top-of-atmosphere (TOA) reflectance, DN is the digital number, RADIO_ADD_OF F SET is the radiometric offset (−1000 for all bands
16
G. Sialelli et al.
for PB≥ 04.00, 0 otherwise), and QU AN T IF ICAT ION _V ALU E = 10,000. For Sentinel-2 L2A data, the conversion is given by Equation 2, where ρBOA is the bottom-of-atmosphere (BOA) reflectance, BOA_ADD_OF F SET is the corresponding offset (−1000 for all bands for PB≥ 04.00, 0 otherwise), and BOA_QU AN T IF ICAT ION _V ALU E = 10,000. All values can be retrieved from the product metadata files. The offset applies to any product with Processing Baseline ≥ 04.00, which should be determined from the PROCESSING_BASELINE field in the metadata rather than the acquisition date: as part of the Collection-1 reprocessing campaign completed in 2024, ESA reprocessed the entire historical archive, so even early acquisitions now carry the offset convention. This procedure applies to products distributed by ESA (e.g., through CDSE). Some third-party platforms (such as GEE’s harmonized collections) apply the offset upstream, in which case it must not be applied a second time.
D
\label {eq:1} \rho _{TOA}=\frac {\text {DN}+\text {RADIO\_ADD\_OFFSET}}{\text {QUANTIFICATION\_VALUE}}
(1)
\label {eq:2} \rho _{BOA}=\frac {\text {DN}+\text {BOA\_ADD\_OFFSET}}{\text {BOA\_QUANTIFICATION\_VALUE}}
(2)
Data Formats for AI4EO Datasets
Table 2 compares the storage formats most commonly used for AI4EO datasets, highlighting trade-offs in random access, cloud readiness, and geospatial metadata support. Table 2: Storage layout and practical properties of data formats relevant for AI4EO datasets. Format
Storage model
Typical payload
Random access
Geo metadata
HDF5
Single binary container with groups, datasets, and internal chunk indices
Heterogeneous N-dim scientific arrays
Contiguous offsets or chunk index
Optional; domainspecific conventions required
NetCDF4
NetCDF data model N-dim variables on HDF5 backend with named dimensions and attributes
HDF5 chunking through the netCDF API
Usually via CF conventions
GeoTIFF
TIFF extended with Regular gridded geospatial keys raster imagery
Strips or tiles; efficient Built in through spatial access requires GeoTIFF tags tiling
COG
GeoTIFF organised Cloud-distributed for partial reads and geospatial rasters overviews
Tile offsets and range Built in through requests; overviews GeoTIFF keys reduce full-scene reads
Zarr
Key-value hierarchy of metadata and encoded chunks
Direct key lookup; v3 sharding groups chunks into larger objects
Chunked Ndim arrays for cloud/distributed processing
Via emerging GeoZarr conventions
Best Practices for EO Maps
E
17
Spatial K-fold block cross-validation
Figure 2 illustrates three splitting strategies and their effect on train–test leakage due to spatial autocorrelation. (a) Random split
(b) Spatial blocks
(c) Blocks + buffer zone
Train/test neighbours share autocorrelated signal → Optimistic evaluation
Folds separated, but cross-fold pairs near boundaries still correlated if blocks are too small → Better, but potential edge leakage
Buffer excludes near-boundary samples from all folds
Train
Test
Fold 1
Fold 2
→ Realistic evaluation
Fold 3
Fold 4
Buffer
Excluded
Fig. 2: Random splits leave train and test points autocorrelated (a). Spatial blocks separate folds geographically but near-boundary points may remain correlated (b). Adding a buffer zone removes this leakage (c).
F
Mitigating patch artifacts
Figure 3 compares the two main strategies for handling border effects when merging patch-level predictions into a seamless map. (a) Overlapping patches + blending
(b) Padding + crop to center
1. Overlapping grid
2. Weighted merge
Patch i+1
Patch i
1. Expand with context
2. Predict → crop
padding
discarded
patch
kept
overlap • Smooth transitions
weight profiles (e.g. cosine)
• No blending — preserves model output
• Blurs predictions in overlap zone
• Uncertainty maps remain faithful
• Complicates uncertainty propagation
• Residual artifacts if receptive field < padding
Fig. 3: Two strategies for mitigating patch artifacts: (a) overlapping patches with weighted blending and (b) padding with center cropping.
G
Candidate grids
Table 3 summarizes the geometric properties of candidate grids for global dataset construction; no single grid satisfies all three desiderata simultaneously.
18
G. Sialelli et al.
Table 3: Comparison of candidate grids for global EO dataset construction. Globally continuous indicates that the grid tiles the entire sphere with each point assigned to a uniquely defined cell and well-defined adjacency across all cell boundaries. ≈ indicates that the property holds approximately but not exactly. , The property holds locally within a single UTM zone, not globally. a Clips at ±85◦ latitude. b Continental boundaries have 50 km overlap. c Via implicit Voronoi tessellation of sample points. d Point-based grid; projection not prescribed. e Grid cells do not align cleanly across the antimeridian. Family
Grid
Equal-area
Conformal
Globally continuous
Geographic
Plate carrée
×
×
✓
Projectionbased
Web Mercator UTM / MGRS Equi7 Grid
× ×, ✓, ≈
✓ ×, ✓, ×
✓a × ×b
DGGS
S2 Geometry H3 HEALPix
≈ ≈ ✓
× × ×
✓ ✓ ✓
Purposebuilt
Major TOM Equal Earth grid
≈c ✓
n/ad ×
✓ ×e