Tetsuya KANAGAWA

The Study of Landslide Detection Based on Deep Learning Trained by Image and Altitude Information

Kazuyoshi Takahashi

The infrastructure facilities such as electricity, gas, and water supply are indispensable for our daily lives. However, Japan frequently experiences disasters such as earthquakes, causing these facilities to be frequently disrupted, particularly in mountainous areas. Early recovery of these facilities is crucial, but initial assessment can be delayed due to the difficulty of visual inspections and poor accessibility to facilities located in mountainous areas. 
To address this issue, this study aims to construct a landslide detection method that can acquire location information for collapsed areas using visible images and altitude information before and after a disaster. The transfer learning method was employed to improve the detection accuracy of landslides, particularly those in shadowy areas and those with remaining vegetation. We first confirmed the effectiveness of altitude information in landslide detection using a single network model and developed a method to improve detection accuracy. We then visualized the misclassification factors using Grad-CAM and revealed that the use of altitude information can result in misclassification of the entire image due to small non-collapse areas. Additionally, we found that the same network model trained with the same data can result in biased feature extraction, leading to difficulty in detecting landslides. Therefore, we employed multi-model detection to reduce the impact of feature bias and evaluated the detection performance using nine networks. Although the detection results included some false positives, we were able to obtain images that captured the shape and location information of landslides by modifying the detection criteria. 
As a result, it was confirmed that landslide conditions can be grasped during a disaster, enabling assessment of infrastructure inspection routes, damage assessment in inaccessible mountainous areas, and estimation of forestry resources damage.