Landslide Detection In SAR Images Based On Deep Learning

The rapid acquisition of landslide disaster information from SAR landslide images is of great significance for emergency rescue.Synthetic Aperture Radar has the outstanding advantage of being able to obtain feature information all day and without being affected by weather,which provides great convenience for the relevant departments to obtain the landslide information in the first time.With the continuous development of this technology,the application of SAR image data in landslide disaster relief has attracted attention,and some excellent results have also been achieved.This paper is devoted to the research of automatic landslide detection and landslide information extraction.Based on the characteristics of SAR image landslide data,deep learning technology is used to automatically detect landslide information,and the application of deep learning technology to landslide disaster detection is completed.The main research content and innovation content of this article are as follows:(1)This paper studies and analyzes the SAR image image characteristics and landslide image characteristics,and obtains the main image performance differences of the non-landslide body and the landslide body.To lay a solid theoretical foundation for the use of deep learning technology to extract SAR landslide image features and design landslide detection algorithms.(2)This paper explores the principle and main flow of SAR image change detection,and focuses on the difference map algorithm of multi-temporal images.Due to the relatively large amount of SAR image data,an algorithm for generating difference maps with less calculation time and good performance is selected,and under this algorithm,the accuracy of landslide extraction is not affected to a certain extent,and the inherent multiplicative noise of SAR images is eliminated.(3)A SAR image landslide detection algorithm based on Gabor Wavelet and LSTM is proposed,and this method is used to realize landslide detection.This algorithm uses a logarithmic ratio operator to generate a difference map based on multi-temporal landslide images,uses Gabor wavelet to extract features from landslide images,and then uses LSTM to learn and classify SAR landslide image features(acquires information on landslides and non-landslides).In addition,the effectiveness of real PALSAR landslide images acquired by JAXA with different classifiers is explored.(4)The study of SAR landslide image features shows that the SAR image landslide area is more distinctive and significant than the landslide area.In this paper,the SAR image landslide detection algorithm based on frequency domain analysis and CWNN is proposed.In this algorithm,frequency domain analysis is adopted to analyze the landslide area significantly,extract more abundant and accurate landslide information,and to a certain extent,suppress the influence of speckle noise in the landslide area on the subsequent deep learning technology to extract landslide information.At the same time,CWNN was selected as the classifier to ensure the translation invariance of its network,which achieved very good results,and also suppressed the influence of noise to a certain extent.The true PALSAR landslide image obtained through JAXA verifies its effectiveness.