Agentic AI for Remote Sensing: Technical Challenges and Research Directions

A team of eight researchers from institutions including MBZUAI, TU Berlin, and others published a position paper on arXiv (2604.24919) arguing that Earth Observation (EO) requires fundamentally different agentic AI architectures than generic models. While foundation models and vision-language models have advanced remote sensing, and agentic AI excels at long-horizon reasoning and tool use, EO workflows operate over georeferenced, multi-modal, temporally structured data. Operations like reprojection, resampling, compositing, and aggregation actively transform underlying state and constrain subsequent analysis, causing errors to propagate silently across steps. The paper identifies that correctness depends not just on internal coherence, but on geospatial consistency, temporally valid comparisons, and physical validity—requirements that break generic agentic models.

The authors propose design principles for EO-native agents centered on structured geospatial state, tool-aware reasoning, verifier-guided execution, and learning objectives aligned with geospatial and physical validity. They outline research directions including EO-specific benchmarks, hybrid supervised and reinforcement learning, constrained self-improvement, and trajectory-level evaluation beyond final-answer accuracy. The paper emphasizes that building reliable geospatial agents requires rethinking agent design around the physical, geospatial, and workflow constraints that govern EO analysis, arguing these challenges are structural rather than incidental.