Development and Validation of an AI System for Identifying Embankment Locations Using SAR Satellite Data and Deep Learning

Abstract

Following the 2021 landslide disaster in Atami, Shizuoka, national demand for embankment management grew, leading to the enforcement of the Embankment Regulation Act in 2023. However, surveying vast areas to identify embankment sites is costly in terms of personnel, time, and expense. This study developed an AI model combining JAXA’s ALOS-2 SAR data with deep learning to extract embankment locations. Using pre- and post-construction SAR intensity images, the model classifies each pixel as “embankment” or “non-embankment,” enabling stable observation even in cloudy or nighttime conditions. Validation in Tottori and Okayama achieved 60-70 % recall at a threshold of 0.1 for sites exceeding 3,000 m². Future work will develop a hybrid model integrating Sentinel-2 optical data to improve accuracy and generalizability.