Ryuji TAKADA

Study on Extraction of Landslide Areas with Transfer Learning Usin Image Information of Pre- and Post- Disaster

Kazuyoshi TAKAHASHI

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, First, as a preliminary experiment, we compared the discrimination results of collapse areas between our originally constructed deep learning model and a model with transfer learning using SqueezeNet. As a result, the recall rate of the collapsed areas and the accuracy were the recall rate of non-collapsed areas was higher for the originally constructed deep learning model. Next, we performed transfer learning using several pre-trained models and compared the discrimination results of collapse areas. The results showed that GoogLeNet had the highest recall rate for collaps areas and accuracy, and VGG-16 had the highest recall rate for non-collapsed areas. Finally, using GoogLeNet, we attempted to detect collapse areas from images with a large area, instead of the patch-by-patch detection we have been doing so far. As a result, the recall rate of collapsed areas was lower than that of patches, but the recall rate of non-collapsed areas was higher. Throughout this study, it was found that the misclassified collapsed areas were mainly those in the shadow areas. Since the present study used only visible image information and failed to detect the collapse sites in the shadow areas, it is considered that other information, such as height, can be used as teacher data for better detection of the collapse sites.