Designing Trustworthy Earth Observation Foundation Models.

A new chapter discusses the critical design principles for developing scalable and trustworthy Foundation Models specifically for Earth Observation (EO). It emphasizes that while foundation models have revolutionized machine learning, their direct transfer to remote sensing (RS) is often suboptimal without domain-specific adaptation, primarily due to the unique physics governing EO data and operational decision constraints. The review synthesizes the current landscape of RSFMs, detailing pretraining objectives, architectural designs, and the crucial requirements for downstream adaptation and trustworthiness. It points out that current evaluation benchmarks are often inconsistent, hindering fair comparison and reliable deployment. The authors advocate for evaluating next-generation RSFMs not just on benchmark accuracy, but also on their ability for modality-aware transfer and their capacity to generate physically plausible representations. This approach is illustrated with case studies like physics-informed masking for harmful algal bloom prediction, underscoring the importance of domain-guided principles for trustworthy EO decisions.