Geospatial Foundation Models

Overview

Geospatial foundation models learn reusable representations of places and Earth processes from satellite imagery, maps, environmental measurements, and other spatial data. Common approaches include self-supervised masked prediction and contrastive learning, multimodal and temporal sensor fusion, and adapting pretrained encoders with lightweight fine-tuning or task-specific decoders. These models support tasks from land-cover mapping and disaster monitoring to urban analysis, forecasting, and geospatial reasoning. A central challenge is reliable transfer across regions, sensors, seasons, and tasks: performance can shift substantially, and shared benchmarks, physical validation, and uncertainty assessment remain important.

Latest papers

134 papers14 papers in the last four weeks, down 12% on the four weeks before. 0.1% of all new papers.

Nov 11, 2026

TopoST: Topology-Conditioned Node Specialization for Urban Spatio-Temporal Foundation Models

Urban spatio-temporal foundation models must share knowledge across diverse urban scenarios while accommodating heterogeneous node dynamics. Our analysis of 20+ datasets spanning diverse urban scenarios reveals recurring local graph structures across datasets alongside substantial node heterogeneity within local subgraphs. Motivated by these observations, we propose TopoST, which combines shared learning over local subgraphs with topology-conditioned node specialization. The key insight is to make node specialization itself transferable through a shared topology-to-parameter mapping. This mapping uses identity-independent topological representations to generate node-specific temporal and spatial dependency parameters, enabling direct specialization for unseen nodes without target-specific embeddings or model updates. Extensive experiments demonstrate that TopoST achieves remarkable zero-shot generalization across unseen cities and tasks, highlighting topology-conditioned node specialization as an effective mechanism for transferable urban spatio-temporal foundation models. Code is available at https://github.com/weikaiyun/TopoST.
Nov 10, 2026

Satellite Pretraining Is Not One Thing: A Matched-Pair Study of Generic and Earth-Observation Foundation Models for Scene Classification

Geospatial foundation models are pretrained on satellite and aerial imagery, but whether such an encoder beats a strong generic one for few-label land-use/land-cover mapping remains unclear, as prior comparisons rarely isolate the pretraining data. We study five matched pairs of Earth-observation (EO) and generic encoders sharing objective and architecture (DINO/SatDINO, DINOv3/DINOv3-SAT, MAE/SatMAE, MAE/Scale-MAE, CLIP/RemoteCLIP) within a benchmark of 14 frozen encoders on five RGB scene-classification datasets (0.06-30 m GSD). With native preprocessing, we evaluate k-NN, linear and attentive probes and image retrieval over six low-label splits, using paired per-image significance tests. EO pretraining is not uniformly helpful: on most datasets and metrics, it significantly improves three pairs (SatDINO, Scale-MAE, RemoteCLIP, the latter evaluated only on unseen datasets) and significantly hurts two (linear probe: DINOv3-SAT by up to 22.5 points, SatMAE by up to 8.8), with opposite outcomes even for two MAEs pretrained on the same fMoW-RGB imagery. These effects are robust to patch-token pooling and attentive probing, concentrate on small man-made objects, and are largest with 1-5% labels. Across encoders, the silhouette score of frozen embeddings predicts k-NN accuracy and retrieval mAP almost perfectly (Spearman ρ≥0.96ρ\geq 0.96 on four of five datasets). Pretraining on satellite imagery is thus not a reliable selection criterion; candidate checkpoints must be validated on target data.
Nov 9, 2026

National-Scale Crop Type Mapping in Morocco Using Google AlphaEarth Foundations Embeddings and the Fields of the World Field Boundary Map

Recent geospatial foundation models have introduced a new paradigm for large-scale Earth observation by providing pretrained embedding representations that can be directly exploited for downstream tasks. In this work, we investigate the use of Google AlphaEarth Foundations embeddings for national-scale crop type mapping in Morocco. The proposed workflow operates directly on publicly available 64-dimensional pretrained embeddings using lightweight downstream classifiers, eliminating the need to process raw satellite imagery during model training and inference. The approach targets three crop categories, namely winter cereals, tree crops, and other crops. For nationwide inference, the Fields of the World global field boundary map is used to constrain predictions to agricultural parcels, enabling efficient large-scale mapping. Experimental evaluation on an independent test set over ten repeated runs achieved an average classification accuracy of (87.17 +/- 0.42%). Nationwide inference over Morocco was completed in approximately five hours, with the dominant computational cost arising from cloud-based retrieval of the embeddings rather than model execution. These results demonstrate the potential of pretrained embedding representations for computationally efficient national-scale crop type mapping.
Nov 8, 2026

Decomposing Earth Embeddings with Sparse Autoencoders

Earth observation foundation models produce global metre-resolution embeddings that underpin work across agriculture, ecology, hydrology, and beyond. Yet these embeddings are opaque: each dimension entangles unrelated physical concepts, which makes them difficult to interpret and their robustness hard to assess. Here, we propose sparse autoencoders (SAEs) as a principled way to decompose Earth embeddings into sparse, approximately monosemantic features over an overcomplete dictionary. Because LLM judges of interpretations do not transfer to satellite pixels, and spatial autocorrelation matters, we introduce a tailored evaluation protocol. It scores any SAE on reconstruction at matched sparsity, alignment with reference concepts, and spatial scale, each against baselines that include an SAE whose decoder is a frozen random dictionary. On this protocol, SAEs reconstruct AlphaEarth and Tessera embeddings better than every baseline at matched sparsity. Their features align with far fewer reference concepts than raw embedding dimensions (on AlphaEarth, 88% align with at most one of 20 concept groups, most with none, against a median of 12 per raw dimension). They also carry more of their variance at fine spatial scales (23 to 32% within a 640 m patch for BatchTopK and Matryoshka, against 9%). An SAE with a frozen random decoder shares both properties, which therefore follow from sparsity, whereas only a learned dictionary reconstructs the embedding accurately: on AlphaEarth, trained SAEs leave eight times less variance unexplained. We test the sparse code on two applications, discovering land-cover sub-concepts and detecting out-of-distribution pixels by area-of-applicability estimation. Both indicate how sparse features could be used, but neither is yet reliable enough for practical use. Together, these results are a step toward Earth observation models whose features can be inspected.
Nov 5, 2026

Towards Label-Efficient Geospatial Foundation Model Adaptation via Positive-Unlabeled Learning for Building Damage Mapping

Geospatial foundation models (GFMs) are reshaping remote sensing by shifting the field from task-specific modeling toward general-purpose geospatial intelligence. Mapping the damage to buildings stands to benefit substantially from this shift, as GFMs pretrained on large-scale Earth observation data provide generalizable geospatial representations that can facilitate effective responses to sudden-onset disasters. Although parameter-efficient fine-tuning (PEFT) facilitates GFM adaptation, parameter efficiency does not necessarily translate into label efficiency. Building damage mapping remains challenging due to scarce pre-existing labeled data or limited time for annotation. When weak supervision is introduced to alleviate the labeling bottleneck, we identify the difficulty of learning sufficiently discriminative representations as a key challenge for label-efficient GFM adaptation. To address this challenge, we propose a Positive-Unlabeled GFM Adaptation (PUMA) framework that requires annotations for only a small number of damaged buildings. More specifically, PUMA integrates non-contrastive weakly supervised representation refinement to mitigate the impact of noisy supervision on representation learning from a gradient perspective, and latent-state-based label disambiguation to enable reversible supervision refinement without requiring class prior estimation. Extensive experiments on disaster benchmark datasets from around the world demonstrated the effectiveness and robustness of PUMA. A real-world application to the 2025 Los Angeles wildfires further highlighted its potential to accurately map damage under highly limited supervision. More broadly, these findings suggest that weakly supervised PEFT offers a label-efficient pathway from GFMs to real-world applications.
Nov 5, 2026

When GeoFMs Meet VHR: Benchmarking Foundation Models on Very High Resolution Earth Observation Imagery

Geospatial foundation models are a widely adopted starting point for remote sensing tasks. Most are pretrained on Sentinel-1/2 and Landsat imagery at a ground sampling distance (GSD) of 10m or coarser. Many applications instead rely on very high resolution (VHR) imagery, where individual buildings, roads, and amenities are resolved. We study which pretraining regimes transfer well to this scale. We introduce VHR-Bench, a benchmark suite of ten datasets at 1.5m GSD or finer spanning five task families, eight application domains, and both satellite and aerial platforms. We evaluate twelve encoders in three pretraining tiers with frozen backbones at two training-data fractions. We find that both pretraining on large natural-image collections and VHR-specific pretraining transfer effectively to VHR tasks, with complementary strengths. SkySense leads when labels are scarce; at the full label fraction the web-pretrained DINOv3 ViT-L edges ahead, with the two close together there. Encoders pretrained only on Sentinel-scale imagery transfer least well under RGB input. Guided by these complementary results, we continue the self-supervised pretraining of DINOv3 ViT-L on 1.02 million held-out VHR tiles and evaluate the resulting encoder without changing the downstream protocol. The continued model gives a clear improvement over the released backbone when labels are scarce and a modest one at the full fraction, at no increase in inference parameters or compute. These findings support a generalize-then-specialize strategy: already strong, broad natural-image pretraining can be further improved by a short VHR continuation stage while preserving its general visual features.
Nov 2, 2026

EO-Jev: Decision Model for Earth Observation

Earth observation (EO) workflows require selecting among known outcomes, estimating confidence, and processing large volumes of imagery. Generative vision-language models can address these tasks, but autoregressive decoding is computationally costly and their confidence estimates are unreliable. Decision models offer an alternative, returning probabilities over candidate answers in a single forward pass. However, their accuracy, calibration, and transfer across EO tasks remain insufficiently understood. We investigate these properties through a diagnostic study of d3-flash (8.4B parameters) across 30 settings from 23 public EO benchmarks. We find that compositional reasoning transfers well and calibration improves over its Qwen3.5-9B backbone on 24 settings, while counting, change quantification, damage grading, and absent-object recognition remain challenging. To address these limitations, we adapt d3-flash through low-rank fine-tuning on public EO training splits, producing EO-Jev. EO-Jev improves significantly over d3-flash on 26 of 30 settings and matches or exceeds specialist references on 20 of 28. It also achieves an average gain of 8.0 points across 12 held-out settings and lower expected calibration error than specialists on 21 of 28 settings. These improvements come at a modest cost of 1.8 points in general decision performance. Our results demonstrate the potential of adapting general decision models into competitive EO generalists while highlighting persistent challenges in quantitative perception, option-order consistency, and calibration under distribution shift.
Oct 31, 2026

Urban functional knowledge transfer through global representations amid uneven data

Urban functional observations are unevenly distributed worldwide, limiting analysis of how places are organised and used. AlphaEarth Foundations (AEF) provides a globally consistent physical representation. Whether this representation can support transfer of functional knowledge to locations with sparse or absent functional observations, and what governs transfer, remain unclear. We introduce POI-aligned Urban Embedding (PURE), which learns physical-functional relationships between POI-derived semantics and AEF. Across five urban analysis tasks covering 142 countries and territories, PURE increases mean R2R^2 from 0.792 to 0.830, with gains across widely different levels of POI availability. When target-city POIs are excluded, country-level alignment retains 87.9% of its performance gain. Cross-country experiments show that transfer is selective rather than universal, with larger gains between physically similar and geographically closer contexts, particularly among pairs within the same continent. Our findings reveal spatially structured urban functional knowledge transfer, offering a route to reduce reliance on dense local observations.
Oct 30, 2026

OCELOT: Planetary-Scale Species Distribution Modeling with Satellite Foundation Embeddings

Large-scale repositories of species observations now contain billions of occurrence records from around the world, but these sightings cluster heavily in human-accessible areas. Existing models use only a small fraction of these records and are not evaluated on extrapolating to remote, unvisited regions--where species distribution models can be used as geographic priors to help classify data from camera traps and acoustic recorders. We scale relative occurrence species distribution models by training a multimodal implicit neural representation network called OCELOT that fuses spatial coordinates with satellite foundation embeddings on two orders of magnitude more observations than previous work: 2 billion GBIF records across 100,000 taxa. Scaling to this regime requires co-designing the data pipeline, taxonomic representation, loss formulation, and model architecture: spatiotemporal thinning and sub-linear cell sampling correct for spatial redundancy and observer bias, while an additive taxonomic readout shares representations across species, genera, and families to improve predictions for rare taxa. To evaluate performance in remote areas, we benchmark geographic extrapolation using the NDCG metric on spatial block holdouts up to 600 km. Satellite foundation embeddings improve extrapolation by 51% over coordinate-only models and outperform classical bioclimatic rasters. The resulting model also matches or outperforms existing baselines on established interpolation benchmarks, demonstrating that optimizing for geographic extrapolation does not compromise in-distribution accuracy.
Oct 29, 2026

NeurEO: dissecting Earth observation embeddings with computational neuroscience

Earth embeddings are compressed representations of rich Earth observation data that enable downstream models to tackle a wide range of complex geospatial tasks. Yet, this increase in performance comes at the cost of decreased interpretability, as the semantic meaning of Earth embeddings is unknown. To address this challenge, we borrow methods from the field of computational neuroscience, which aim to interpret activity in (biological) neurons. Here, we adapt and apply these methods to assign meaning to AlphaEarth embeddings. We analyse embedding features both as stand-alone data, equivalent to spontaneous neural activity, and in relation to land cover and elevation correlates, equivalent to stimulus-response neural activity. We demonstrate correlational and multi-scale structure in the embedding features. Embedding features can display highly specific and directional tuning to land cover and elevation, such as a `North vs South woodland edge' response. When we remove the main trends of combined land cover responses (like water versus land tuning), we discover fine-grained representational details with increased effective dimensionality. Together, our analyses start to unravel what these Earth embeddings represent, which is critical for developing trustworthy downstream applications.
Oct 27, 2026

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

Earth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations remains limited. Geospatial foundation models (FMs) generate high-dimensional embeddings encoding complex spatio-temporal land-surface properties, but their utility and interpretability for agricultural applications remain insufficiently understood. This study presents a representation-agnostic framework evaluating a 64-dimensional AlphaEarth embedding space against Sentinel-2 reflectance, NDVI phenology, and harmonic features across geometric separability, temporal dynamics, spatial robustness, and cross-regional generalisation using field-scale data from the Conterminous United States (CONUS; 2017-2023) and Northern India (2021). Traditional features achieve strong linear separability (Fisher ratio), while embeddings provide superior distribution-aware separation - Bhattacharyya and Mahalanobis distances up to an order of magnitude higher for spectrally similar pairs (e.g. Corn/Soybeans: 3.9 vs. 0.10). Embeddings uniquely capture multi-year crop rotation dynamics, yielding a separability ratio of 1.784, versus a maximum of 1.003 for all engineered features. Under spatially explicit validation, embeddings exhibit lower residual spatial autocorrelation of errors (mean Moran's I≈0.17I \approx 0.17), reducing reliance on geographic proximity. However, harmonic features show stronger cross-continental transferability (US →\to India Macro F1: 0.493 vs. 0.326 for embeddings), with NDVI showing a similar advantage. The combined feature space yields balanced performance. These findings reveal a trade-off between representational richness and cross-region transferability, providing a framework for feature selection in large-scale crop classification.
Oct 26, 2026

Accuracy and efficiency of compact foundation models for surface-water mapping during floods

Pretrained Earth observation models may reduce the effort needed to map surface water, but their value depends on the alternative against which they are compared. We evaluated Prithvi-EO-2.0-Tiny using the 446 hand-labeled Sentinel-2 chips from 11 flood events in Sen1Floods11. A fine-tuned model, a frozen-encoder model, and an identical randomly initialized architecture were compared with a similarly sized U-Net, a larger Prithvi model, and spectral-index thresholds. Experiments varied the number of training chips, excluded events from model fitting, and measured computational cost and prediction reliability. With 252 training chips, U-Net achieved a water intersection-over-union (IoU) of 0.856 on the official test set, compared with 0.833 for fine-tuned Prithvi-Tiny. Pretraining improved the tiny transformer over random initialization by a reported 0.046 IoU with 25 training chips and 0.023 with 252 chips, but did not reverse the U-Net ranking. Across 11 event-exclusion evaluations, mean IoU was 0.768 for fine-tuned Prithvi-Tiny and 0.777 for U-Net; the paired interval included zero. Fine-tuned Prithvi-Tiny processed a 512x512-pixel chip in 16.1 ms on a GPU and 258 ms on a laptop CPU, compared with 30.6 and 665 ms for U-Net. Errors were concentrated near water boundaries, and high predictive entropy identified many, but not all, erroneous pixels. These results support a specific role for compact pretrained models: lower measured computational demand with competitive accuracy, rather than consistent superiority. Event-dependent performance, shared hyperparameter selection, and the distinction between observed water and newly inundated land delimit the conclusions.
Oct 23, 2026

Reuse, Don't Retrain: Efficient, Modality-Versatile Geospatial Foundation Models through Feature Distillation

Geospatial foundation models (GeoFMs) are costly to pretrain, yet model families often repeat this process across backbone scales and sensing modalities. We investigate feature distillation as a more efficient way to reuse existing representations. First, we distill frozen CROMA and TerraMind teachers into smaller Vision Transformers through patch-level feature matching. Across nine PANGAEA tasks, distilled TerraMind Tiny, Small, and Base outperform their scale-matched released counterparts by 2.76, 3.02, and 1.87 mean IoU points, respectively; each also matches or exceeds the next-larger released model. Distilled CROMA-Base gains 2.05 points. TerraMind-Base required only 43 A6000 GPU-hours, approximately 0.5% of the reported cost of independent pretraining. Second, we introduce TEDI, an unpaired multi-expert approach combining TerraMind for multispectral imagery and DINOv3 for high-resolution RGB. Separate patch embeddings route unpaired inputs into a shared Transformer trunk, removing the need for co-registered data. TEDI-Large achieves the highest single-model mean in our PANGAEA-9 comparison (62.51 mIoU), outperforming both teachers overall while retaining their modality-specific strengths. TEDI-Base reaches 61.75 mIoU with 72% fewer parameters. These results show that distillation improves the GeoFM accuracy-compute frontier while consolidating complementary experts into a single versatile backbone.
Oct 20, 2026

Towards Consistent Geospatial Representations: Aligning One Encoder to Many Foundation Models

Geospatial foundation models (GFMs) differ widely in pretraining objective, modality and resolution, and none leads across benchmarks. We ask whether their representations are complementary, and whether a single encoder can inherit that complementarity. We introduce Pangaea-Latents, an extension of the Pangaea benchmark providing frozen embeddings of 55 GFMs over its 1111 datasets, along with reproduced baselines, and find GFM embeddings to be neither redundant nor disjoint. We then propose GeoMixture, a preliminary framework to align a ViT to several GFMs per dataset through a multi-representation alignment (mREPA) objective. Our encoder ties the best GFM in average rank across nine Pangaea datasets, trading peak performance for cross-task consistency.
Oct 14, 2026

PiCo: Physics-Informed Concept Probing for Geo-Foundation Models

Geo-foundation models (GFMs) are evaluated by downstream predictive performance. This reveals little about which physical properties their embeddings are sensitive to or which are used in predictions. This limits developers' ability to test hypotheses about model behaviour and users' ability to select models for new tasks. We propose Physics-Informed Concept Probing (PiCo), a post-hoc framework for explaining GFM predictions. PiCo identifies concept vectors in GFM embedding space using synthetic multispectral satellite observations from a 3D radiative transfer model coupled with a game engine. These vectors characterise GFM sensitivities to physical properties, informing domain-based hypotheses about downstream performance. Projecting embeddings onto concept vectors produces low-dimensional representations for quantifying concept importance in predictions. We evaluate PiCo on three GFMs with different input modalities using seven physical concepts covering forest leaf optics and canopy structure. In ground-cover estimation, RGB-only inputs reduce dry-matter sensitivity and disproportionately degrade predictions for dry vegetation and bare ground, as anticipated by PiCo's sensitivity analysis. For hierarchical forest type and tree species classification, concept representations retain 70.0--93.8% of full-embedding accuracy. Different GFMs capture physical differences between classes, such as canopy structure between broadleaf forest and cleared ground and leaf-water-content between oak and beech. Yet they differ in the strength of these contrasts and the concepts they use for classification. PiCo provides an interpretable basis for hypothesis testing and model assessment, with potential applicability beyond Earth observation wherever physical models enable controlled simulation and pretrained embeddings are accessible for probing. Code is available at https://github.com/yihshe/PiCo.git.
Oct 14, 2026

MARS - A foundational map auto-regressor

Map generation tasks feature extensive non-structural vectorized data (e.g., points, polylines, and polygons) and thus pose significant challenges to common pixel-wise generative models. Conventional approaches use multiple stages, first segmenting these features at the pixel level and then performing vectorized post-processing, with errors and complexity compounding at each stage. Motivated by the recent success of auto-regressive language modeling, we propose the first map foundation model, named Map Auto-Regressor (MARS), that is capable of generating both multi-polyline road networks and polygon buildings in a unified manner. For training MARS, we collected to our knowledge the largest multi-class map extraction dataset totaling 3.4M examples, which we call MAP-3M. Across four road and building datasets, MARS outperforms or matches the performance of multistage baselines. Additionally, we develop a "Chat with MARS" capability that enables interactive human-in-the-loop map generation and correction, supported by the auto-regressive nature of our end-to-end approach. We release our MAP-3M dataset and project demo page at (1) https://huggingface.co/datasets/bag-lab/MAP-3M and (2) https://huggingface.co/spaces/bag-lab/MARS, respectively.
Oct 13, 2026

MajorTOM:Elliot Pretrain, A Multi-Sensor, Multi-Temporal Pretraining Dataset for Earth Observation

Pretraining generalist Earth observation foundation models requires data that combine several sensors, varied temporal sampling and, increasingly, text, yet open datasets rarely provide all three in one consistent, easily accessible form. We introduce MajorTOM:Elliot Pretrain, a collection of 279,166 tiles of 10 km drawn from the Major-TOM grid by clustering AlphaEarth embeddings. Each tile stacks Sentinel-2 L1C, Landsat-8/9, Sentinel-1 RTC, Copernicus DEM and ESA WorldCover on a shared 1056x1056-pixel, 10 m grid, complemented by cloud masks, OpenStreetMap entities and ERA5 weather. The tiles are partitioned into three temporal regimes, Monotemporal (250,000 tiles, one acquisition), Monthly (12,500 tiles, twelve acquisitions spread over the annual cycle) and Burst (16,666 tiles, six acquisitions within 30 days), and are released as 6.3 TB of Cloud Optimized GeoTIFFs indexed by TACO v3 Parquet metadata, a standard that allows cloud streaming, efficient data access, and filtering before any pixel is read. A companion library derives sensor-tagged fact sheets, captions and 16 families of vision-language tasks from these layers at load time. We test our design choices in short pretraining runs of OlmoEarth-Base. First, a 25% subset of Major-TOM Core-S2L2A curated with AlphaEarth embeddings matches or exceeds the full corpus on 8 of 9 benchmarks. Second, on Elliot-Pretrain, adding Landsat-8/9 improves or matches Sentinel-2 only pretraining on every benchmark. Third, using DEM and WorldCover as prediction targets during pretraining further improves 8 of 9 benchmarks, while giving the same maps to the encoder as inputs degrades 5 of 9. Fourth, Sentinel-1 helps on 7 of 9 benchmarks, particularly on benchmarks with radar inputs. Fifth, rule-based captions match a small language model at no cost. The dataset is openly available at https://source.coop/major-tom/elliot-pretrain.
Oct 11, 2026

Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation

Google Earth AI is a family of geospatial AI models, together with an agent that orchestrates them. We present this agent, Google Earth AI's Geospatial Reasoning agent (GRA), a Gemini-powered research system that combines geospatial data sources, foundation models, and tools to answer questions about our planet. GRA organizes these into modular domain skills and pairs them with upfront geospatial planning to decompose complex queries into multi-stage workflows. Evaluating such systems is hard: real-world problems admit multiple valid workflows, require data discovery across large catalogs, and rely on evolving ground truth. We introduce two benchmarks and release their development splits. GeoExpertQueries poses complex analytical problems grounded in practitioner use cases and refined by geospatial experts; each is paired with a consensus golden workflow from three independent experts and scored by a human-validated, tool-agnostic autorater. GeoKTools consists of templated queries with deterministic reference solvers that systematically vary the number of required tool types. We also use two external benchmarks, CBGB and UnivEARTH; the latter we find to be saturated. To isolate the contribution of GRA's tools and its agentic orchestration, we compare four configurations, each with Gemini 3.1 Pro and 3.8 Flash: a standalone LLM, the LLM with basic tools, the LLM with GRA's skills and tools but without its upfront planning and execution guidance, and GRA itself. With at most basic tools, an advanced model recovers much of the answer on templated queries but satisfies only a small fraction of the expert workflows on open-ended problems; GRA performs well on both. The largest step comes from GRA's skills and tools; on open-ended problems, its planning and execution guidance add a further gain with Gemini 3.8 Flash, letting Flash match or exceed Pro at lower cost.
Oct 9, 2026

Shortcut Learning in Geospatial Foundation Model Embeddings for Cross-Disaster Damage Assessment

Geospatial foundation models can support disaster damage assessment, but high accuracy on familiar disasters may hide shortcut learning: prediction from event or scene characteristics rather than evidence that transfers to a new disaster. We study embeddings from TerraMind, Prithvi-EO-2.0, DOFA, and DINOv3 on the multi-disaster datasets xView2 and BRIGHT. Holding out an entire disaster, instead of randomly splitting image tiles, reduces mean balanced accuracy from 0.807 to 0.638 on xView2 and from 0.818 to 0.568 on BRIGHT. Four sets of tests examine what information supports prediction: event and acquisition context, before/after image inputs, building count, and correct image pairing. Pre-event images remain predictive, and a classifier given only the annotated number of buildings reaches balanced accuracy of 0.629 and 0.653. Combining before/after features performs best on average on xView2, but does not improve over pre-only features on BRIGHT. Shuffling image pairs within the same disaster and damage class causes modest losses. Motivated by these findings, we test whether feature transformations learned only from training disasters improve generalization to unseen disasters. These transformations improve performance, with supervised contrastive learning performing best on xView2 and PCA on BRIGHT. Our results highlight the importance of evaluation on unseen domains for assessing model robustness and show that simple feature-level transformations can improve generalization. Code and experiments can be found at https://github.com/alishibli97/gfm-shortcut-learning.
Oct 9, 2026

PlacePulse: Learning Transferable Place Representations from Structural Geographic Data

Geospatial machine learning often depends on task-specific feature engineering over large collections of sociodemographic, business, and environmental variables. Geospatial foundation models offer reusable representations of place, but leading approaches typically encode dynamic signals, such as mobility and search activity, or visual signals from satellite imagery. Few models represent the static structural composition of places (who lives and works there, which businesses operate there, and which risks they face) as transferable embeddings at fine spatial resolution. We introduce PlacePulse, a structural geospatial foundation model that compresses 572 attributes, expanded to 2,288 inputs through H3 neighborhood aggregates, into 256-dimensional unit-norm embeddings for 1,055,537 H3 resolution-7 cells in the United States. We also introduce a reusable two-axis benchmark that separates intrinsic representation quality from downstream predictive utility across multiple targets and spatial supports. Intrinsically, the embeddings reconstruct standardized inputs with mean R2=0.853R^2 = 0.853, use 251 of 256 approximately decorrelated dimensions, and show strong spatial organization, with Moran's I=0.93I = 0.93 on the leading principal component and autocorrelation persisting beyond the input radius. Extrinsically, across 74 held-out targets from 12 external sources and four spatial supports, PlacePulse outperforms a dimensionality-matched feature-selection baseline on targets unrelated to the training inputs and is competitive with Google's Population Dynamics Foundation Model, leading on health, economic, election, economic-mobility and social-vulnerability outcomes, with relative strengths that vary by domain. These findings position structural composition as a distinct and complementary modality for geospatial foundation models, not simply a fallback where behavioral or imagery data are unavailable.
Oct 5, 2026

FairRSFM: A Biome-Aware Benchmark and Debiasing Framework for Remote Sensing Foundation Models

Remote sensing foundation models (RSFMs) are commonly evaluated using aggregate metrics, which can hide systematic performance disparities across ecological regions. We introduce FairRSFM, a biome-aware benchmark for evaluating ecological group robustness in RSFMs. FairRSFM maps georeferenced samples from 14 terrestrial biome classes into six ecologically meaningful macro-groups and evaluates models under a unified frozen-backbone evaluation protocol. The benchmark covers four downstream datasets: m-EuroSAT, m-BigEarthNet, m-SA-Crop-Type, and MMEarth20K with Dynamic World label maps. Using Prithvi-EO-2.0, SatMAE, and DOFA across three random seeds, we show that aggregate performance consistently masks biome-dependent disparities across architectures and tasks. For example, Prithvi-EO-2.0 reaches 90.98% overall macro-F1 on m-EuroSAT but a mean worst-group score of only 83.72%, while m-SA-Crop-Type drops from 27.30% overall mIoU to 18.47% in the Xeric and Mineralogical group. We further evaluate Biome-Orthogonal Linear Probing (BOLP), Dynamic Biome Reweighting (DBR), and GroupDRO as complementary mitigation baselines. Their effectiveness is model- and task-dependent; for example, BOLP improves Prithvi-EO-2.0 worst-group F1@opt on m-BigEarthNet from 46.12% to 50.27% without updating the RSFM backbone. FairRSFM provides a reusable protocol for diagnosing and mitigating ecological robustness gaps in remote sensing foundation models. Code and datasets are available at: https://github.com/aminurhossain/FairRSFM.
Oct 5, 2026

Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health

The efficacy of traditional disease prediction is limited by spatial gaps and temporal lags, which impact the timing and targets of resource deployments. Outbreaks escalate undetected, chronic disease burdens are quantified years later, and at-risk populations in data-sparse regions remain unaddressed. Planetary geospatial foundation models complement existing epidemiological workflows to provide operational improvements, encoding multimodal search, mobility, and environmental signals into generalizable place representations. As illustrations of this complementarity, we present independent global health case studies of Google Earth AI's Population Dynamics Foundation Model (PDFM) -- a foundation model for geospatial inference -- across four domains (vaccine-preventable, communicable, noncommunicable, maternal mental health), five tasks (spatial extrapolation, interpolation/nowcasting, probabilistic forecasting, prospective forecasting, risk stratification), and four countries (USA, Canada, Mexico, and the Democratic Republic of the Congo). Across these case studies, PDFM addresses critical surveillance gaps across domains: improving US-Canada border MMR vaccination coverage predictions by capturing cross-border behavioral spillovers domestic models miss; nowcasting cardiovascular disease to accelerate data availability; enhancing short-term municipal Mexican dengue forecasts for timely outbreak vector control; improving forecasts of cholera hotspots; and adding a transferable signal to individual-level postpartum-depression risk prediction in US states the model had never seen, while not replacing individual socioeconomic data or closing demographic screening gaps. Together, these results showcase capabilities of geospatial foundation models for public health surveillance.
Sep 28, 2026

When local gains fail to transfer: Frozen Earth-observation embeddings across wildfires

Frozen Earth-observation embeddings are judged almost entirely by spatially blocked cross-validation inside one study region. We show that this number does not predict accuracy in a new region; we show why; and we show the one setting in which such a model does keep working, using a protocol that needs only a linear probe and labels one already has. The testbed is wildfire, with Copernicus burned-area maps of six fires in Greece and Spain and descriptors from the year before each fire, comparing TESSERA and AlphaEarth with ESA WorldCover classes and annual Sentinel-2 index summaries. Inside a fire, the embeddings identify the burned land 0.05 to 0.13 ROC AUC better than the index summaries, and repeated fold allocations, spatial buffers, a block bootstrap, and gradient-boosted trees leave that margin unchanged. On a fire in another region, they lose 0.15 to 0.18 AUC, and the index summaries lose 0.06, so the three end within a few hundredths of each other. The representation is not the cause. Eight labelled blocks from the new region restore the embedding advantage and give a higher AUC than 59,000 labelled pixels from other regions, and the weight vector fitted in one region is nearly orthogonal to the vector fitted in the others, so the part that carries across regions is small and low-dimensional. Forecasting within a region is a different matter. Fitted on a fire that burned in 2023 and applied to a fire twelve kilometres away that burned in 2024, where nothing used postdates the target fire, TESSERA reaches 0.772 AUC and loses 0.04 against a classifier fitted inside the 2024 fire, while classifiers fitted in other regions lose 0.09 to 0.18. A region with one mapped fire can therefore forecast susceptibility for later fires there; a region without one cannot borrow a model from elsewhere, and every evaluation of a frozen embedding should report a held-out region.
Sep 26, 2026

Reuse or Relearn? A Spectral View of Earth Observation Foundation Models

Foundation models are rarely used as generic, frozen feature extractors; instead, they are fine-tuned for the target downstream application. This practice is particularly prevalent in Earth observation (EO), and it raises a question that downstream accuracy alone cannot answer: does fine-tuning reuse the pretrained representation, or does it relearn a new one? We study this with spectral diagnostics that compare a model before and after adaptation, quantifying how well its dominant singular subspaces are preserved, how broadly the weight update is distributed, and how large it is. Using natural image models such as CLIP and DINO as a reference, we find that, under the evaluated fine-tuning settings, EO models undergo far larger, higher-rank updates and retain much less of their pretrained structure, so their downstream performance is often obtained with substantial changes to the pretrained weight structure. The diagnostics further provide insight into how cheaply a model can be adapted: where the pretrained subspaces are preserved, adapting a small fraction of the parameters can match full fine-tuning, and where they are not, it can fall behind. More broadly, foundation models, and EO foundation models in particular, should be assessed not only by benchmark accuracy, but also by how reusable their pretrained representation is.
Sep 26, 2026

GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation

Paired synthetic aperture radar (SAR) and electro-optical (EO) imagery is increasingly available across sensors, resolutions, and geographic regions. Yet existing SAR-to-EO image translation (SET) methods are typically trained on a single, limited-scale dataset, producing models specialized to particular sensing conditions. We introduce GeoSET, the first generalist model for SET, built around a single pretrained parent that is adapted to downstream datasets under a common protocol. We curate over 3 million high-quality SAR--EO pairs from a collection of more than 10 million SAR observations, spanning diverse sensors, spatial resolutions, and ground sampling distances. To bridge the modality gap between SAR observations and a pretrained image generator, we develop a speckle-robust SAR encoder and pretrain the conditional generator on this heterogeneous corpus. The resulting parent supports efficient adaptation across downstream datasets through low-rank adaptation (LoRA), updating only 0.60% of the generator parameters and requiring approximately one hour per dataset. Across six downstream benchmarks, GeoSET achieves state-of-the-art results in FID and DISTS with full fine-tuning or LoRA, demonstrating effective transfer across heterogeneous SAR-EO domains.
Sep 24, 2026

AlphaEarth distinguishes cities but compresses urban variation

Cities differ in built form, land cover and development history, complicating comparison across places and time. Satellite foundation models map Earth's surface onto common numerical representations. Yet the tasks and targets used to shape them typically do not focus on cities: globally consistent labels for urban function do not exist, and many datasets - especially land cover and land use classifications - collapse the built environment into few classes. Here we audit the representation, focusing on AlphaEarth but with broader applicability to other Earth embeddings, by probing the geometry and geography of embeddings for 1,000 urban areas in 162 countries. We find that cities occupy a shifted but overlapping region on the hypersphere, 62.7° from the global mean direction, and continent and climate predict 24.3% of variation among the mean directions of urban centres in excluded countries. Inside cities, degrees of urbanisation carry 8.9% of the variation, and what they leave holds shared directions whose local orientation varies, not one universal axis of urbanisation. Retained variation is itself unequal: dispersion within urban centres is 14.1% greater per standard deviation of national development, even after adjusting for population, land area and continent. Further controls suggest cities in developing countries present less contrast in vegetation and texture, and dispersion follows that contrast: full adjustment for it leaves at most 6.4% of the gradient. Annually, a city's representation moves nearly eight times more than redrawing its own pixels explains, and contracts where the 2022 loss of Sentinel-1B removed a pass direction. AlphaEarth's representations therefore support comparison across regions, while the differences between its annual layers are not yet validated for comparison over time.
Sep 24, 2026

Recoverable Geographic Location Information in Earth-Observation Embeddings

Earth-observation (EO) foundation models provide reusable embeddings, yet downstream task accuracy does not reveal whether these representations encode geographic information, which may be beneficial for location-aware applications but potentially detrimental when representations invariant to geographic location are desired. We therefore evaluate the geographic coordinate robustness of Tessera v1, Tessera v1.1, and AlphaEarth by testing whether coordinates can be predicted from the embedding representations using 284 quality-verified European solar farms from 2024. We assessed geographic information content information through the association between cosine and geodesic distances and through prediction of projected coordinates in EPSG:3035. Embeddings from all three EO foundation models contain recoverable geographic information. All prediction models significantly outperform training-range uniform random sampling baselines, with AlphaEarth exhibiting the strongest distance association and lowest mean geodesic error. Both Tessera variants also yielded higher geographic distance correlations than the Sentinel-2 controls. These findings motivate geographic information content as an additional criterion for auditing EO foundation models.
Sep 23, 2026

Geospatial embeddings detect old-growth forests but buffered spatial validation narrows their advantage over Sentinel features

Old-growth forests develop over centuries under minimal anthropogenic disturbance, producing structurally complex and biodiverse stands. In Europe, protecting them requires mapping that is accurate for individual forest parcels yet deployable continent-wide. Geospatial foundation model (GFM) embeddings enable label-scarce land classification, but their value for old-growth detection remains unknown. Here, we map old-growth forests across 211,893 ha of Romania's Southern Carpathians, a beech-spruce landscape typical of the Alpine Biogeographic Region. We construct high-confidence, expert-informed reference labels for old-growth and non-old-growth parcels. We add AlphaEarth, TESSERA v2 and Sentinel-1/2 features to a common baseline of topographic and human-access predictors, then compare them under spatially blocked validation with and without 10 km train-test buffers to limit residual autocorrelation. With buffering, GFM and Sentinel-1/2 predictors increase precision-recall AUC by 0.21-0.25 [95% CIs: 0.15-0.34] relative to baseline, indicating spectral data contain a spatially robust old-growth signal. With a PR-AUC of 0.84 [0.79-0.88], TESSERA outperforms Sentinel-1/2 (+0.08 [+0.05 to +0.11]) and AlphaEarth (+0.08 [+0.04 to +0.12]) under unbuffered spatial validation. At a 10 km buffer, however, this advantage narrows to +0.04 [-0.01 to +0.11] and +0.03 [-0.04 to +0.10], intervals consistent with no difference. At 10 m resolution, convolutional neural networks add no benefit over pixel-based XGBoost. Comparisons with four national- and continental-scale products show the importance of non-old-growth labels, and reveal 81% agreement between our predictions and a field-calibrated map. We conclude that buffered spatial validation is vital when transferring old-growth detection models to unseen landscapes, and provide our labels and predictions for future work.
Sep 20, 2026

RSPDBench: Benchmarking Vision Foundation Models on Earth Observation Tasks Under Physically Grounded Remote-Sensing Product Degradations

Vision foundation models for Earth observation (EO) are commonly evaluated on clean benchmarks, but operational EO products can already contain spatial, radiometric, alignment, noise, and harmonization defects before reaching the model. Generic image corruptions and broad domain shifts do not isolate these product-level failures. We introduce RSPDBench, a physically grounded remote-sensing-product degradation benchmark that evaluates seven foundation-model entries and two supervised baselines on five EO datasets under primitive degradations and compound product chains. Its degradation parameters follow published Sentinel-2 and Landsat~8 specifications and are checked against Sentinel-2's paired 10 m10\,\mathrm{m} and 20 m20\,\mathrm{m} near-infrared bands. We evaluate each model under its own native protocol and measure robustness as the drop from its own clean baseline. Our analysis reveals that degradation sensitivity is strongly structured: (i) the resolution-conditioned encoder is the most robust to resolution loss yet the most sensitive to radiometric gain on satellite products, (ii) the same physical defect can hurt one model while helping another, and (iii) model rankings under resolution loss and radiometric gain are largely reversed. Compound chains add up to 3838 percentage points of drop beyond their strongest component. On m-BigEarthNet, retraining the probe on degraded views removes about 75% of the noise and harmonization losses, but resolution loss still costs about 20% of the clean score. EO robustness must therefore be measured against the structured defects of remote-sensing products, not by clean accuracy or generic perturbations alone. Code: https://github.com/Tanjim131/RSPDBench
Sep 20, 2026

Transferability and operational reliability of a Prithvi crop classification foundation model under phenological and geographic shift across three continents

Fine-tuned geospatial foundation models (GeoFMs) pretrained on large satellite archives have been shown to improve crop classification accuracy and geographic transferability. However, their operational performance beyond the training distribution remains poorly characterized. We evaluated the out-of-distribution performance of a widely adopted GeoFM [Prithvi-EO-2.0] across 37 events in 12 countries on three continents and validated against regional reference products. Results indicated that the mean overall accuracy (OA) declined from 0.65 in the United States to 0.40 in Europe. Beyond accuracy metrics, we assessed five key aspects of model performance: whether model confidence indicates signal failure, sensitivity to observation windows, the effect of coarsening class schemes, and robustness to both band loss and cloud- and shadow-contamination. Accuracy collapsed when the observation window misaligned with local crop phenology, while deterministic confidence remained high. Expected calibration error increased for seven of eight paired events, and 12-51% of each affected scene was confidently mislabeled at near-zero precision. Monte Carlo dropout entropy registered the shift in all eight, indicating that much of the apparent cross-continent decline reflected phenological misalignment rather than spatial transfer. Two adjustments recovered accuracy without retraining. Consolidating 13 classes into 10, based on the model's dominant confusions, raised the mean OA by 8.4 percentage points. Compressing the window toward near-real-time use preserved accuracy across a 45- to 90-day plateau, peaking near 75 days, though arms tighter than 30 days fell about 0.11 below that plateau. Fine-tuned crop GeoFMs therefore transfer usefully only where observation windows match local growing seasons. We translate these findings into operational guidance for the reliable deployment of the released model.