Bi-temporal Remote Sensing Image Change Detection Based On Deep Learning

Land cover and its changes significantly affect the Earth’s biosphere,hydrosphere and atmosphere.In recent years,the change detection algorithms based on convolutional neural networks have achieved superior performance,but most of them have problems such as being bound by the framework of semantic segmentation algorithm and the network is too large for practical application.In this thesis,we focus on the problem of change detection of bi-temporal very-high resolution optical remote sensing image,and devote ourselves to improving the accuracy and practicality of change detection methods based on convolutional neural networks.The main work and results of this paper are as follows.(1)An effective remote sensing image change detection network based on siamese network architecture,Fork Net,is proposed.The major difference between Fork Net and previous work is that its feature extraction phase includes not only a deep feature extraction backbone network,but also a cross-resolution attention module and a contextual information diffusion path.This makes the features output from the feature extraction phase of Fork Net have similar consistent strong semantics at each layer,enhancing the image feature representation discriminative power.To better train Fork Net,we design a new loss,Multiscale Tversky Loss(Ms T Loss),which can generate losses on sub-regions of different scales so that change regions of different sizes can be treated equally,resulting in finer change detection results.The proposed network is trained with a hybrid loss function of Ms T Loss and Focal Loss,and Fork Net achieves the best detection performance on two challenging datasets.(2)An efficient and lightweight remote sensing image change detection network based on difference-aware and representation refinement,Lite-DRNet,is designed.lightweight network is essential for practical applications,and Lite-DRNet is suitable for change detection of very-high resolution remote sensing images in resource-constrained scenarios.Lite-DRNet contains two backbone networks with shared parameters to extract multi-level strong semantic feature maps.Unlike existing change detection methods with large network architectures,we add a difference-aware and representation refinement module after the backbone network to generate an enhanced multi-level image feature representation by leveraging the difference information embedded in the bi-temporal image feature representation.The difference-aware and representation refinement module can be easily integrated into other change detection networks.At the same time,we redesign the difference discrimination network so that it can both fully reuse the feature maps and efficiently combine information from multi-scale feature maps,and keep the parameters and computational effort efficient.The proposed approach is validated on two challenging datasets,and Lite-DRNet achieves leading performance compared to other lightweight networks based on large network architectures,with lower number of parameters and computation.Compared with large networks,Lite-DRNet is lightweight enough but still achieves competitive performance.