Scalable and Trustworthy Earth Observation Foundation Models

The Shift to Domain-Specific Foundation Models

Foundation models (FMs) have revolutionized machine learning by enabling general-purpose models to be adapted to various downstream tasks. However, the authors argue that applying these models to Earth observation (EO) requires a fundamental shift in design. Unlike natural images, which are typically RGB-based and human-centric, EO data are georeferenced, sensor-dependent, and physically measured. Treating RSFMs as merely larger versions of standard vision models ignores the critical physical context—such as atmospheric correction, spectral band structure, and temporal revisit schedules—that defines the utility of satellite data.

Designing for Physical and Operational Constraints

The paper emphasizes that successful RSFMs must integrate domain-specific design principles. This includes moving away from generic RGB pretraining toward architectures that can handle multimodal inputs (e.g., optical, SAR, and thermal data) and varying spatial/spectral resolutions. The authors highlight that pretraining objectives—such as masked image modeling or contrastive learning—must be carefully chosen to respect the physical nature of the data. For instance, random masking might discard critical diagnostic wavelengths, whereas physics-informed spectral masking can force the model to learn meaningful environmental relationships.

The Evaluation Challenge

A central theme of the paper is the current crisis in model evaluation. Because EO tasks range from discrete classification (e.g., land cover) to continuous regression (e.g., chlorophyll concentration), a single benchmark score is insufficient. The authors note that inconsistent evaluation protocols hinder fair comparisons and reliable deployment. They advocate for a shift toward 'capability-specific' evaluation, where models are tested on their ability to perform modality-aware transfer, maintain physical consistency, and provide uncertainty quantification for operational decision-making.

As RSFMs become increasingly integrated into environmental monitoring and decision support systems, the risk of relying on 'black-box' models that lack physical grounding grows. By aligning model architecture and evaluation with the realities of Earth observation, researchers can develop more trustworthy tools for critical applications like harmful algal bloom prediction and disaster response. This framework provides a roadmap for moving from high-accuracy benchmarks to scientifically robust, actionable AI.