SAR backscatter changes can identify flood-related areas, but extracted boundaries may be incomplete. This study combines multitemporal NISAR observations with the Segment Anything Model (SAM) to map flood-related changes following the flooding of 27–28 August 2026 on Japan’s Noto Peninsula.
Three pre-event and one post-event NISAR L2 GCOV dual-polarization images were processed on a common 10 m grid. HH and HV changes were calculated against a pre-event median reference. HH changes were smoothed, corrected for a global offset, and thresholded using robust statistics. Slope, permanent-water, and minimum patch-size constraints yielded 490 candidate patches. Manual review selected six SAR seeds covering 3.5032 km².
Seed-derived point and box prompts guided SAM segmentation of a post-event false-color image combining HH, HV, and their dB difference. Three masks were generated per patch, and one was manually selected based on seed overlap and image boundaries. Joint screening used seed coverage, precision relative to the seed, and intersection over union. Masks for three patches passed, adding 0.9109 km². Retaining all six SAR cores produced a final area of 4.4141 km², a 26.0% increase. Different screening rules produced similar total areas but different local extents.
Independent verification matched stratified sample locations with official photographs and road-inundation records. Of 120 samples with clear flood evidence, 80 fell within SAR cores, 20 within SAM additions, and 20 outside the map. These records support a flood-related interpretation at some added locations. Water may have receded before the satellite overpass, so outside locations cannot automatically be treated as detection omissions.
The workflow provides a semi-automatic mapping approach for flat farmland near rivers. Results are limited to six selected seeds and their accepted additions and do not represent maximum inundation extent.