Abstract
Remote sensing systems now collect large amounts of data from satellites, drones, and aerial platforms. However, turning this data into useful insights still depends on slow, expert-driven processes, often built around separate tools, manual scripts, and task-specific steps. Key challenges remain: combining inputs from different sensors with varying resolutions and spectral bands, analyzing time-series data for change detection or damage assessment, and turning user questions into clear and reliable tool-based steps. To address these issues, we need AI systems that can connect multi-modal visual understanding with structured planning. This requires strong visual models trained on global-scale data, vision-language systems that can handle multi-resolution and time-series imagery, and tool-based agents that can reason about spatial layouts, distances, areas, and other geospatial measurements. Equally important are benchmarks that evaluate how well these agents understand user instructions, select the right tools, and carry out each step with accuracy and traceability. Together, these directions move us closer to interactive, transparent, and scalable analysis in Earth observation.
For more info, please follow this link.