GeoSight v2: strengthening disaster impact assessment with coordinate referencing, inpainting, and similarity models

Post-disaster imagery collected by communities and field teams provides critical visual evidence for assessing damage and guiding emergency responses. However, the geocoordinates embedded in these images often capture the camera’s location rather than the position of the depicted building, introducing spatial misalignment that reduces mapping accuracy and weakens machine learning–based decision support. To address this challenge, we present an enhanced version of GeoSight, a geolocation refinement framework that integrates three key components: (i) coordinate referencing to constrain candidate search areas, (ii) building detection with inpainting to isolate structures and remove unrelated objects such as vehicles and debris, and (iii) perceptual similarity analysis using DreamSim to match query images against a georeferenced database. Using NOAA’s 2023 Norman, Oklahoma tornado building-damage imagery, the enhanced framework improves retrieval accuracy from 53% to 67% (Top-1) and 77% to 84% (Top-3), while reducing average geolocation errors by 26.1 m, with maximum corrections up to 46.5 m. By combining coordinate data, object-level preprocessing, and perceptual similarity, this framework provides a scalable and generalizable approach for refining the geolocation of community-driven disaster imagery, enabling more accurate and timely damage mapping for disaster response and recovery.