Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain

Researchers argue that farmland remote sensing segmentation cannot rely on a single current image because crop appearances change over time and can be visually ambiguous.

Their proposed FarmSeeker agent detects uncertain regions, explains the ambiguity, and asks for extra spatio-temporal information to refine segmentation.

They also introduce GSFS-Bench, a global high-resolution benchmark for reasoning and query-based evaluation in farmland segmentation.

Experiments show more stable performance than prior methods, suggesting a shift from static image analysis to active decision-making for tougher scenes.