Study On Convolution Neural Network For Change Detection On Multi-Source Remote Sensing Images

Remote sensing technology has become an efficient approach in the survey of national land resource utilization by the right of acquiring multitemporal images over the same geographical area.The ability for a single sensor to acquire the target scene periodically are tightly bound by the satellite revisiting period and imaging quality,while collaborative observation using multi-sensor provides high-frequency,multimodal remote sensing image,which fulfills the massive high-quality data in land-cover change monitoring.Oriented to the application of land resources monitoring in the era of big data,the deep convolution neural networks are explored and introduced in this thesis to realize the change detection using homologous and multi-source high resolution remote sensing images.To handling the multi-scale feature learning and the optimizing of convolution,the learning of dilated convolution,innovative design of siamese network architecture and supervision training method have been studied.The main research contents of this paper are as follows.(1)To handling the issues of complicated background and multi-scale features in remote sensing change detection,an innovative convolution neural network with multiscale feature(DSCNH)has been proposed.Firstly,Hybrid Convolutional Feature Extraction Module(HCFEM)is designed based on dilated convolution and pointwise convolution operation.To achieve the spatial aggregation on low dimensional embedding,the hierarchical parallel strategy is adopted on the feature representation phase.Then,the discrimination module of the differentiation characteristics is included according to the Siamese neural network.After that,a change detection model named Deep Siamese Convolutional Network with HCFEM(DSCNH)is proposed.Moreover,the bootstrap method and object based uncertainty analysis are introduced in the training process post-processing phase respectively,which improves the detection performance of the model.To demonstrate the ability of DSCNH for the change detection task,four kind of datasets were carried out.Experimental results showed that the proposed model achieved the best performance.(2)An innovative Mixed Interleaved Group Convolution Network(MIGCNet)has been proposed to modify the convolution operations on fine-grained level.Firstly,the mixed interleaved group convolution(MIGC)is proposed by mixing multiple convolution kernel sizes in one convolution operation,which contributes to the multiscale convolutional features learning without any expanding.Then,the differentiated feature extraction and the feature fusion are proposed to extract the shallow difference information on two channel and obtain the deep ensemble features.Four kinds of datasets,including multi-source and homologous images,were carried out to demonstrate the superior performance of MIGCNet.(3)To handling the hard training issues of deep neural network,the multi-loss supervision training strategy has been proposed in the thesis.In this method,different parameter optimizers are utilized to train different parts of MIGCNet and the intermediate layers receive feedback gradient directly from change results,which reduce the gradient vanishing during the training processing of deep network.Experiments on four kinds of datasets showed that the multi-loss supervision training method can improve the performance of MIGCNet on differences identification.This article has 47 figures,18 tables and 132 references.