17:15 〜 18:45
[HTT15-P03] Google Earth Engineと深層学習を用いた日本の表層崩壊のマッピング
キーワード：表層崩壊、深層学習、グーグルアースエンジン、画像の分類とセグメンテーション
Shallow landslides represent significant hazards in Japan, particularly in regions like Shizuoka and Fukuoka Prefectures, emphasizing the urgent need for accurate mapping techniques. This study introduces a novel approach integrating the Google Earth Engine with image deep learning techniques to effectively map shallow landslide areas. Remote sensing data, such as multispectral satellite imagery from Sentinel-2 and Planet, provide unprecedented capabilities for comprehensive Earth surface monitoring. We systematically evaluate various deep learning models for this purpose and discuss the selection of the optimal model. The results show that the Residual Network (ResNet) demonstrates rapid and accurate classification, achieving a training accuracy of approximately 99%. Conversely, the Medical Transformer (MedT) exhibits optimal performance in shallow landslide detection, with an F1 score of 0.78 and a Dice coefficient of 0.53, followed by DeeplabV3+ and Unet++. Our findings contribute to detecting shallow landslide areas, aiding in prioritizing and planning risk mitigation measures to reduce potential impacts on human life and infrastructure.