Research On Multi-temporal Remote Sensing Target Change Detection And Classification Based On Deep Learning

As a kind of earth observation data,remote sensing image contains a large amount of information in space.How to extract the information is very important to the application of remote sensing image.Therefore,the recognition of remote sensing image and the detection of remote sensing target change have become the research hotspot at home and abroad,and the application field is more and more extensive.Because of the characteristics of multi band and high resolution of remote sensing image,the feature extraction of remote sensing image is more complex.There are some problems such as the information feature extraction is not obvious,the change detection accuracy is not high,and the information utilization of remote sensing image is insufficient.In order to improve the accuracy of remote sensing image information extraction and change detection,this paper applies the capsule network model and the Siamese network framework to the field of remote sensing image classification prediction and change detection,and improves the capsule network and change detection method.The main work of this paper is as follows:(1)Two improved capsule network methods are proposed and applied to remote sensing image classification and prediction.This paper there are two improvements to the capsule network,one is to improve the activation function of the capsule network,the other is to improve the structure of the capsule network.Among them,the improvement of activation function can enlarge the smaller vector and avoid the orthogonality of feature vector,resulting in zero result;by adding pooling layer and feature fusion layer in capsule network,it can adapt to different sizes of remote sensing image input and extract more features.The experimental results show that the performance of the capsule network with improved activation function and network structure is better than that of other baseline methods.(2)A change detection method of twin residual neural network is proposed.For remote sensing image change detection.For the problems of "pseudo change",this paper designs the process of change detection,such as remote sensing image preprocessing,super-pixel segmentation and merging,change feature analysis and change feature extraction.First of all,remote sensing image preprocessing and super-pixel segmentation and merging can effectively reduce the impact of noise such as "pseudo change" on change detection accuracy.Secondly,through the analysis of change characteristics,the sample selection results are obtained.Finally,the twin residual network is used for change detection to improve the accuracy of change detection.Through the test of experiment one and experiment two,the results show that the improved method improves the overall performance index,which shows that the change detection method in this paper is effective.