Detection of Landslide Candidate Interference Fringes in DInSAR Imagery using Deep Learning

Jyoko KAMIYAMA, Tomoyuki NORO, Masayuki SAKAGAMI, Yamato SUZUKI, Kazuo YOSHIKAWA, Shuhei HIKOSAKA and Ikushi HIRATA

Interferometric synthetic aperture radar (InSAR) is an effective technique for monitoring the risk of large-scale sediment movements because it can broadly and routinely observe the extent of landslides. To detect interference fringes with the possibility of landslide movements from differential interferograms, it is common for experts to interpret these fringes considering the effects of water vapor as well as the topography and other factors. Increasing the accuracy of detecting landslides is an important issue in the usage of InSAR. Convolutional Neural Networks (CNNs) that enable image recognition with sufficient accuracy have recently been developed. To efficiently detect landslide candidate interference fringes, this study evaluated the effectiveness of introducing a CNN model to detect the interference fringes representing landslide movements using similar processes as experts techniques. As a result, the CNN model was able to detect landslide candidate interference fringes that had been detected by experts with a recall of approximately 90%.

2018/1 416-425