Landslide Classification from Synthetic Aperture Radar Images Using Convolutional Neural Network with Multichannel Information
S. Mabu, Y. Nakayama, T. Kuremoto
J. Signal Processing., 2020, 24, 61-73.
DOI : 10.2299/jsp.24.61

Abstract: Detection of disaster-stricken areas using synthetic aperture radar (SAR) images is important in countries and regions with heavy rain and earthquakes. Although it is important to immediately find disaster-stricken areas when a disaster occurs, it takes time to read SAR images and also needs experience and expertise. Therefore, machine learning, especially deep learning, is expected to be applied to the classification of disaster-stricken areas. Classification using deep learning is often executed on patch images of local areas. However, patch-based classification would miss information on the surrounding areas such as topographic features. In this study, a convolutional neural network (CNN) is applied to the classification of SAR images using the following techniques. When making the images input to a CNN, two multichannel image generation methods, i.e, a zero-padding method and map-concatenation method, are used, where the target areas to be classified and their surrounding areas are combined to form multichannel images. In the experiments, the zero-padding method and map-concatenation method are evaluated by the classification performance of SAR images that cover the northern Kyushu area in Japan, where large-scale landslides due to heavy rain occurred in 2017. Through the experiments, we clarify the appropriate CNN structures with multichannel information for landslide classification.