Anomaly Detection Using Convolutional Adversarial Autoencoder and One-class SVM for Landslide Area Detection from Synthetic Aperture Radar Images

Abstract

In Japan, typhoons often approach and pass from July to September every year because of the effects of westerly winds and high atmospheric pressures. Landslide disasters frequently occur due to heavy rains caused by typhoons, which leads to major accidents related to transportation and human life. Until now, observation of the disaster areas, e.g., landslides, has been mainly conducted by aircraft [1]. However, in recent years, remote sensing using satellite images has attracted attention as a method for observing a wide area [2]. In satellite remote sensing, a sensor is mounted on an artificial satellite to observe the ground surface of the earth, and the obtained image is analyzed. Therefore, we can see the damages without going directly to the disaster areas. Typical satellite images include optical images that are obtained from sunlight reflection, and Synthetic Aperture Radar (SAR) images that are obtained by sensors that emit microwaves to the ground surface. It is easy for human eyes to interpret optical images, but they cannot be observed at night or in bad weather. On the other hand, SAR can observe the surface of the earth regardless of time and weather; thus, SAR images are useful for rapid rescue activities at night and in bad weather conditions. However, it is difficult for human eyes to interpret SAR images, unlike optical images. Therefore, many methods that analyze SAR images have been proposed to detect disaster areas rapidly [3,4], and machine learning techniques, especially deep learning, have also been applied to landslide area detection [5].