Authors: Hsiang-Chun Chen*, The Department of Geography, National Taiwan Normal University, Kuo-Chen Chang, The Department of Geography, National Taiwan Normal University
Topics: Remote Sensing
Keywords: Landslide, deep learning, Image interpretation
Session Type: Paper
Presentation File: No File Uploaded
Man-made improper development combined with fragile geological conditions and typhoons or earthquakes usually lead to severe disasters in mountainous or slope areas in Taiwan. Therefore, the government attaches great importance to the monitoring of landslide. However, when we use non-deep learning methods to identify different images, it often takes a lot of manpower and time to update the model parameters, which makes it impossible to use the same model efficiently on different images. The concept of deep learning proposed by Hinton in 2006 is based on a large number of parameters that can be automatically adjusted. In the process of learning to identify the target object, the parameters are constantly fine-tuned so that the computer can successfully identify the same target object when encountering different images. The deep learning method was confirmed in 2012, which can effectively solve the problem that non-deep learning methods require human intervention when encountering different images. This study uses the image recognition methods commonly used in deep learning—Faster R-CNN and YOLOv2—to automatically identify the landslide in Taiwan through the Formosa satellite imagery. And we comparing the conditions of landslide of different spatial characteristics, images of different seasons, different spectral combinations of input image, etc., to judged which method has better performance. Finally, we find out which kind of deep learning algorithm is more suitable for identifying the landslide, so as to effectively solve the obstacle of this major disaster prevention mechanism.