A STUDY ON THE AUTOMATIC DETECTION OF LANDSLIDE FROM AERIAL PHOTOGRAPHS USING UNet AND UNet++

Abstract

 Japan is covered with mountains and hills, and it is a problem that there are many disasters such as landslides caused by earthquakes and torrential rains. At present, the judgment of the sediment disaster section is done by visual interpretation work from the aerial photograph taken after the disaster. However, this work requires a long period of visual observation by an experienced person, which places a heavy burden on the worker. In order to improve the efficiency of this work, the study on automatic detection of sediment moving parts by deep learning using aerial photographs has been advanced in recent years. In this study, in order to improve the accuracy of automatic detection of sediment moving parts by UNet, the authors applied UNet++ as an image segmentation model, which has the feature of suppressing false detection and is expected to improve the accuracy of sediment moving part detection.