The Deployment Gap of Geospatial Foundation Models
An unsigned editorial in Nature Machine Intelligence noted that a single survey covering the three-year window from mid-2021 to mid-2024 documented 58 remote sensing vision foundation models. Prithvi, Clay, AnySat, SkySense, and SatMAE are among them, each reporting impressive performance across land cover classification, change detection, and environmental monitoring, and each leveraging large-scale self-supervised pre-training to claim generalization over diverse global data distributions (Butsko et al., 2025). Benchmarks like PANGAEA and GEO-Bench now span resolutions from 0.1 to 30 meters, multiple sensor types including optical and SAR, and tasks from segmentation to regression. The architecture problem is largely solved. Self-supervised pre-training on heterogeneous Earth observation imagery works.
What the models have not done, in most cases, is run in production. Despite benchmark results that warrant serious attention, the deployment of geospatial foundation models in operational settings is both challenging and rare, because standardized evaluation tasks frequently fail to capture data heterogeneity, resource constraints, and application-specific requirements that determine whether a model is actually usable (Butsko et al., 2025). The distance between a model that performs well on GeoBench and a model that produces outputs a global crop-mapping system can consume every week is not an architectural distance, it is a pipeline distance.
Geospatial data does not arrive pre-processed. An operational Earth observation pipeline routinely draws from Sentinel-1, Sentinel-2, digital elevation models, and weather datasets (Butsko et al., 2025), each at different resolutions, in different coordinate reference systems, with different spectral band configurations and temporal cadences. ESA’s Copernicus archives alone held over 66 petabytes of satellite data as of 2024, and that figure grows continuously. Prithvi, trained on NASA’s Harmonized Landsat-Sentinel data, ingested six spectral bands at 30-meter resolution delivered as Cloud-Optimized GeoTIFF, with preprocessing that included atmospheric correction, cloud masking, and spatial co-registration. An operational pipeline that delivers four Sentinel-2 bands at 10 meters cannot straightforwardly feed that model. Band mismatch, resolution mismatch, and coordinate inconsistency each require resolution not at inference time, but in the pipeline architecture that governs what reaches inference at all.
Pretraining data composition turns out to exert as much influence on downstream performance as architectural choices. Most existing benchmarks are geographically skewed toward North America and Europe, and the pre-training datasets that feed the models often reflect the same bias, directly limiting generalizability outside those regions (Marsocci et al., 2025). The broader challenge is that geospatial data is non-stationary across geography, sensor, and time in ways that most pre-training pipelines do not fully account for: ESA’s Copernicus corpus is large, but breadth and representativeness are not the same thing (Marsocci et al., 2025). The decisions made during data curation and sampling, before a single training step, shape what the model is capable of handling in deployment.
Ground truth is the second constraint. Labeled geospatial data is spatially and temporally imbalanced, expensive to produce, and concentrated along geographic and class lines that reflect annotation cost and domain expertise rather than the distributions that matter for real applications (Liu et al., 2025). A model trained on labels collected before the target deployment window faces temporal shift relative to the conditions it will actually see. Benchmark accuracy does not transfer automatically when the label distribution in the field differs from the one the benchmark controlled for.
Then there is compute. Operational systems typically run on limited computational resources, often without access to GPU nodes, while benchmark evaluations routinely assume otherwise. That gap is the first problem operational practitioners have to solve before a model can run at all (Butsko et al., 2025). A model that achieves state-of-the-art flood mapping performance on a GPU cluster may not be deployable in the monitoring infrastructure where flood maps are actually needed.
The tooling layer is beginning to close this distance. STAC and Cloud-Optimized GeoTIFF established a queryable, cloud-native pattern for Earth observation cataloging and data delivery. GeoParquet extended comparable conventions into the vector domain. TerraTorch is built on PyTorch Lightning with specific adaptations for the heterogeneity of Earth observation data, including varying spectral bands, resolutions, and the specialized domain knowledge the field requires, and provides a common fine-tuning and benchmarking interface across multiple foundation models. Its 1.0 release natively supports the Prithvi family, Clay, SatMAE, Satlas, DOFA, and others, with parameter-efficient fine-tuning via LoRA and ViT-Adapter. PANGAEA provides a unified evaluation framework across eleven benchmark datasets and multiple sensor modalities, designed to be extended as new models and datasets emerge. These tools do not dissolve the pipeline problems, but they standardize where the work happens, which is the precondition for addressing those problems systematically rather than per-deployment.
The WorldCereal global crop-mapping system is the most detailed public record of what it actually takes to move from benchmark-capable model to operational output. The challenges it surfaces are the routine conditions of any Earth observation pipeline operating at scale: data heterogeneity, temporal shift in labels, resource constraints that preclude GPU access, and the absence of standard protocols for moving a foundation model from benchmark to production. The model contributed to the system; the pipeline governed what it could deliver.
The issue is not which architecture posts the best number on the next benchmark. It is what the operational integration layer needs to look like for those models to reach the monitoring systems, mapping pipelines, and decision workflows where they are actually needed.
Clairvoyint helps expert teams turn analytical methodology into reusable, evidence-grounded agents. If your team relies on complex analysis that depends on expert judgment, fragmented tools, and defensible outputs, let’s talk.