Geospatial Foundation Models

• GFMs are large-scale, pre-trained models that extract and generalize high-level representations from diverse Earth observation data for tasks like scene classification and change detection.
• They employ advanced methodologies such as Vision Transformers, masked image modeling, and teacher–student continual pretraining to effectively handle multi-sensor inputs.
• GFMs optimize computational efficiency while ensuring robust transfer learning across modalities and regions, supporting sustainable geospatial AI applications.

Geospatial Foundation Models (GFMs) are large-scale, pre-trained models specifically designed to extract, generalize, and transfer high-level representations from Earth observation and geospatial data, enabling a wide range of downstream applications including scene classification, semantic segmentation, change detection, and environmental modeling. GFMs leverage architectural paradigms from natural language processing and computer vision, notably transformers and masked image modeling, but are distinct in their focus on integrating remote sensing (RS), multi-sensor, and other complex geospatial modalities. Their rapid evolution capitalizes on advances in dataset diversity, multi-modal fusion, scalable training, and computational efficiency, positioning them as central tools for modern geospatial artificial intelligence and planetary monitoring.

1. Principles and Methodologies of GFM Construction

GFMs are architectured upon modern backbone models such as Vision Transformers (ViT), Swin Transformers, and their multi-scale or multi-expert derivatives. The pretraining pipeline commonly adopts self-supervised learning (SSL) objectives—especially masked image modeling (MIM), generative contrastive frameworks, and masked autoencoding (MAE)—to ingest vast volumes of heterogeneous, unlabeled EO data (Tsaris et al., 17 Apr 2024, Jia et al., 10 Mar 2025).

Model design is tailored for geospatial characteristics:

• Multi-spectral/multi-modal input handling: Models like Prithvi process 6+ bands, while msGFM introduces sensor-specific embeddings and shared encoders for unified cross-modal representation (Blumenstiel et al., 4 Mar 2024, Han et al., 1 Apr 2024).
• Teacher–student continual pretraining: This paradigm leverages robust, frozen representations from large-scale natural-image models (e.g., ImageNet-22k pretraining) and adapts students via joint reconstruction and feature alignment losses:

$L = L_{\mathrm{MIM}} + L_{\mathrm{feat}}$

where $L_{\mathrm{MIM}} = \|O_k - G_k\|_1 / N$ (masked region L1 loss), and $$L_{\mathrm{feat}} = - (P(f_L^S)/\|P(f_L^S)\|_2) \cdot (f_L^T/\|f_L^T\|_2)$$ (feature distillation via cosine similarity) (Mendieta et al., 2023).

• Diffusion-based SSL: Generative geospatial diffusion models (e.g., SatDiFuser) demonstrate that hierarchical, noise-conditioned representations—fused via global/local/mixture-of-experts (MoE) strategies—can rival or surpass strictly discriminative backbones in downstream tasks (Jia et al., 10 Mar 2025).
• Parameter-efficient adaptation (e.g., DEFLECT): Embedding deflection mechanisms finely disentangle spectral and spatial cues, enabling adaptation of RGB-trained models to more general multispectral settings with only 5–10× fewer additional parameters than conventional low-rank adaptation (Thoreau et al., 12 Mar 2025).

The selection and composition of pretraining data is also a critical design factor; balanced, globally representative sampling (such as stratified-continental/biome or uniform-at-random) yields superior downstream performance over regionally-clustered datasets across tasks and architectures (Purohit et al., 21 Jan 2025).

2. Dataset Design and Multisensor Integration

The effectiveness and generality of GFMs are directly influenced by the diversity, representativeness, and modality coverage of their pretraining datasets:

• Diversity and feature entropy: GeoPile aggregates imagery from sources including NAIP (1 m GSD), RSD46-WHU, MLRSNet, RESISC45, and PatternNet, optimizing for higher entropy (4.6 vs. 3.9 for Sentinel-2 alone) and spatial/spectral heterogeneity (Mendieta et al., 2023).
• Phenology-informed, seasonal sampling: Datasets such as SSL4Eco use MODIS-derived EVI breakpoints (Greenup, Maturity, Senescence, Dormancy) to sample Sentinel-2 imagery per local phenological cycles, vastly improving ecological and seasonal feature learning (Plekhanova et al., 25 Apr 2025).
• Multisensor and cross-sensor harmonization: msGFM demonstrates a masked image modeling framework that employs sensor-specific patch embeddings, a shared transformer, and cross-sensor reconstruction losses:

$I'_{(i)} = D_i(\mathrm{En}(f_j(I_{(j)})))$

where $f_j$ and $D_i$ are the embedding and decoding layers for modalities $j$ and $i$ respectively (Han et al., 1 Apr 2024).

Such design minimizes domain bias (e.g., from natural or urban scenes in high-activity areas), supports multi-temporal and multi-modal fusion, and enables robust transfer learning across unobserved regions, sensor types, and ecological contexts.

3. Performance, Evaluation Protocols, and Benchmarking

Evaluating GFMs involves systematic, reproducible protocols and the use of diverse, representative benchmarks to capture real-world task complexity:

• Unified evaluation frameworks: PANGAEA establishes inclusive benchmarks covering wildfire mapping, marine pollution, agriculture, disaster damage, biomass estimation, across multiple sensors (optical, multi-spectral, SAR) and geographic/temporal domains (Marsocci et al., 5 Dec 2024).
• Downstream tasks: Typical benchmarks involve semantic segmentation (mean IoU), land-cover classification (F1, mAP), change detection (F1, accuracy), and regression (RMSE, $R^2$ for aboveground biomass, climate variables, or asset metrics) (Muszynski et al., 28 Jun 2024, Ghamisi et al., 30 May 2025).
• Comparison to supervised and prior self-supervised baselines: GFMs that include domain- and task-relevant pretraining (e.g., SSL4EO-L with Landsat for Landsat-Bench tasks) demonstrate 4–5% higher OA/mAP versus ImageNet-pretrained models (Corley et al., 10 Jun 2025). For remote sensing retrieval, the Prithvi model achieves mAP 97.62% (BigEarthNet-43) and 44.51% (ForestNet-12), outperforming RGB-only models (Blumenstiel et al., 4 Mar 2024).

Under label scarcity (low-data regimes), GFMs pre-trained on spectrally diverse or high-resolution imagery maintain performance advantages, while in full supervision regimes, traditional architectures (e.g., UNet) remain competitive in certain simpler or regionally-biased scenarios (Marsocci et al., 5 Dec 2024).

4. Computational Efficiency, Scalability, and Environmental Impact

Scaling of GFMs addresses both computational requirements and environmental considerations:

• Model scaling effects: Scaling ViT models from 100M to 3B parameters yields up to 30% improvement in top-1 scene classification accuracy, with further scaling to 15B parameters explored for future adaptability (Tsaris et al., 17 Apr 2024).
• Resource and carbon impact: Multi-objective continual pretraining (combining MIM and feature distillation) reduces GPU hours and CO₂ emissions by nearly an order of magnitude compared to training from scratch (e.g., 93.3 V100 GPU hours and 13.3 kg CO₂ for GFM, versus 768 hours and 109.44 kg CO₂ for SatMAE) (Mendieta et al., 2023).
• Model distillation and deployment: Frameworks such as InstaGeo compress model size by up to 8× via teacher–student distillation. The distilled student optimizes task accuracy plus a distillation loss (commonly KL-divergence on logits), often yielding negligible loss in mIoU (e.g., –0.73 pp for flood, +1.79 pp for locusts) (Yusuf et al., 7 Oct 2025).
• Automated optimization: Toolkits like TerraTorch integrate Bayesian hyperparameter optimization (via Optuna) and support no-code, modular fine-tuning workflows, reducing expertise and time barriers for geospatial research (Gomes et al., 26 Mar 2025).

These methods facilitate large-scale model training on supercomputing clusters (e.g., Oak Ridge’s Frontier), as well as rapid, low-carbon deployment and benchmarking through web-based map applications and open-source pipelines.

5. Multi-modal, Multi-source, and Application-driven GFMs

Modern GFMs increasingly target integration across imagery, vector, and contextual datasets:

• Multimodal fusion: GeoLink unifies RS imagery and OSM data via joint contrastive pretraining and spatially aware cross-modal transformers. Fusion objectives include an intermodal InfoNCE loss and spatial-consistency constraints to ensure geographic object-patch alignment:

$$\mathcal{L}_{\mathrm{cont}} = - \frac{1}{2N} \left[ \sum_{i} \frac{\exp(\mathrm{sim}(z_G^i, z_I^i)/\tau)}{\sum_j \exp(\mathrm{sim}(z_G^i, z_I^j)/\tau)} + \cdots \right]$$

$$\mathcal{L}_{\mathrm{cst}} = \frac{1}{N} \sum_{i} \frac{1}{L^m_V} \sum_j \| \epsilon_{OR}^m[i,j] - \sigma_V^m[i,j] \|^2$$

(Bai et al., 30 Sep 2025).

• Extendability and generalizability: Tools such as TerraTorch and PANGAEA enable reproducible evaluation across new domains, support for new modalities, and comparative performance reporting, furthering the standardization of GFM research (Gomes et al., 26 Mar 2025, Marsocci et al., 5 Dec 2024).
• Societal relevance: GFMs are applied in SDG-grounded benchmarks for asset wealth, health prediction, environmental hazard mapping, and biodiversity tracking, with performance measured not only by accuracy but also energy efficiency and operational carbon footprint (Ghamisi et al., 30 May 2025).

6. Security, Privacy, and Ethical Considerations

GFMs, due to their scale and data diversity, introduce unique privacy and security risks:

• Lifecycle vulnerabilities: Risks include data memorization of private spatial attributes, identity linkage across modalities, centralized weight leaks, prompt-based attacks, and poisoned feedback in RLHF/RLAIF loops (Rao et al., 2023).
• Privacy-preserving strategies: Solutions encompass geomasking, K-anonymity, differential privacy (with guarantees such as $Pr[M(D) ∈ S] ≤ e^{\epsilon} Pr[M(D') ∈ S]$), federated learning to decentralize sensitive data, and robust prompt engineering (Rao et al., 2023).
• Policy and benchmarking: The field is moving toward standardized protocols for privacy/robustness evaluation and calls for interdisciplinary collaboration in regulatory design.

7. Limitations, Outlook, and Future Directions

Current challenges and avenues for future improvement in GFMs include:

• Adapting to additional modalities: Expanding beyond RGB to SAR, hyperspectral, LiDAR, and temporal series to address wider range of EO applications (Han et al., 1 Apr 2024).
• Domain adaptation and data distribution sensitivity: Architectural sensitivity to pretraining data distribution necessitates robust, diverse, and well-sampled global datasets to avoid bias and improve transferability (Purohit et al., 21 Jan 2025).
• Composable and scalable approaches: Feature-level ensembling and knowledge distillation provide practical pathways for scaling GFMs. Ensembles such as Hiera_Prithvi_500M can exceed larger monolithic GFMs while also allowing subsequent student distillation for operational deployments (Chuc, 25 Jun 2025).
• Environmental, social, and ethical considerations: There is increased emphasis on carbon impact, transferability, and real-world utility, moving model evaluation from purely technical benchmarks to societal and policy-driven outcomes (Ghamisi et al., 30 May 2025, Yusuf et al., 7 Oct 2025).

GFMs are poised to serve as foundational tools for next-generation, sustainable, and robust geospatial AI, driving progress in planetary understanding, environmental management, and global development challenges.