Research On Remote Sensing Image Landslide Information Extraction Algorithm Based On Deep Learning

Landslides are a type of global distribution and high frequency natural disaster,which has the characteristics of sudden,multiple and mass.Landslides cause great damage to the safety of human life and property and environmental resources.Therefore,it is of great significance for disaster reduction and relief to extract location information from the landslide in time.With the development of remote sensing satellite technology,higher resolution remote sensing image can be obtained by satelliete,however,there are still more challenges in landslide identification with high-resolution remote sensing images.For the traditional image-based classification method only uses single spectral information,which makes the feature information too simple,the model can’t dig deeper information,and the traditional machine learning classifier makes the accuracy of landslide recognition low.Based on the combination of pixel-based classification method and deep learning,depth features of high-resolution remote sensing images are fully utilized to extract landslide information,a remote sensing recognition method is proposed on the basis of the depthwise separable convolutional neural network of the miultimixed spectral characteristics.The main research work includes:(1)This thesis discusses the feature combination of remote sensing image classification on pixel-based classification of remote sensing image and the improvement of neural network for landslide characteristics.Based on the characteristics of high resolution images of landslides,the CNN(Convolutional Neural Networks)us improved and a DSCNN(deep separable convolutional neural network)model based on the characteristics of landslides is obtained.In addition,NIR(near infrared spectroscopy)band,NDVI index features,and the texture information is extracted to compensate for the uniqueness of spectral information and enhance the classification effect.Images with homologous sensors(the spatial information of images before and after the disaster is the same)but not in the same direction are used.By combining the pre-landslide remote sensing images with the post-landslide images and other texture information,the multifusion band values reflecting the changing characteristics of the landslide are determined,which can effectively overcome the tedious steps of the traditional hierarchical classification method and enhance the classification effect.(2)A landslide recognition method is proposed on the basis of the depthwise separable convolutional neural network of the miultimixed spectral characteristics in this thesis.Based on the pixel-based classification,multiple mixed bands are used as eigenvalue,the improved depth separable convolution neural network is used as classifier for supervised learning to identify the landslide,so as to finally complete the identification of landslide.In this thesis,three kinds of characteristics and four kinds of classifiers are compared,and parameters such as precision and f-measure are used for performance evaluation.The results show that,compared with other algorithms,the recognition accuracy of this algorithm is higher,which can meet the practical application requirements,and is of great significance for disaster prevention and relief by using high-resolution remote sensing images.