Comparison between result of transfer learning using pre-trained networks on landslide areas detection

• Published in: Journal of the Japan society of photogrammetry and remote sensing Vol. 60; no. 6; pp. 350 - 353
• Main Authors: TAKADA, Ryuji, KANAGAWA, Tetsuya, OKA, Shigeaki, KAWAMURA, Naoaki, ONITSUKA, Shunichi, TAKAHASHI, Kazuyoshi
• Format: Journal Article
• Language: English; Japanese
• Published: Tokyo Japan Society of Photogrammetry and Remote Sensing 2021; Japan Science and Technology Agency
• Subjects: collapse area detection; Deep learning; Disaster recovery; Disasters; Image classification; Landslides; Machine learning; Mountainous areas; Recall; Transfer learning; vegetation
• ISSN: 0285-5844; 1883-9061
• DOI: 10.4287/jsprs.60.350

A quick assessment of landslide damage in mountainous areas after a disaster occurs is important for planning of the disaster recovery action. For this assessment, deep learning AI is thought to be an effective method to grasp quickly the state after a disaster. The deep learning for image classification, however, needs a large amount of training and test data. To overcome this problem, a transfer learning is thought to be effective, especially, when much data is not available. In this paper, we compared between each result of four major pretrained CNN architectures (AlexNet, GoogLeNet, VGG-16, SqueezeNet) to be transfer learned, using pre- and post-disaster visible image information. At the result, GoogLeNet showed best collaspe recall rate 81.4%, and accuracy 85.7%. Also, VGG-16 showed the best non-collapse recall rate 93.1%. The remarkable point of this comparison is that collapse areas, which were classified as non-collapse areas by all of four model, belonged to ground involved large amount of sand and clods.