Remote Sensing Image Semantic Segmentation Based On Deep Learning And Domain Adaptation

Semantic segmentation of remote sensing images aims to classify all pixels in remote sensing images,which has important practical value in the fields of resource planning,ecological monitoring and disaster assessment.With the rapid development of deep learning,convolutional networks and Transformer networks have been able to automatically learn meaningful feature representations from images,thus becoming the preferred semantic segmentation methods nowadays.However,deep learning models still have the following shortcomings: 1)the training of deep learning highly relies on large-scale high-quality datasets,however,the cost of labeling a high-resolution remote sensing image is too high,thus limiting the wide application of deep learning in the field of remote sensing? 2)remote sensing image datasets suffer from various phenomena such as heterogeneity same spectrum and homogeneity different spectrum,which greatly reduce the generalization performance of deep learning models.To address the above issues,thesis optimizes and improves the technology of remote sensing image semantic segmentation based on deep learning from two perspectives: neural network model architecture and domain adaptive strategy,and the main work and contributions are summarized as follows.(1)A semantic segmentation network Conv Former is proposed based on a hierarchical multi-headed self-attention mechanism and multi-scale feature fusion.In the feature extraction phase,the network chops and embeds images in a sliding window manner,and then captures the long-range dependencies of images at different scales using a multi-layer self-attention module,while learning the position of image blocks using a lightweight feedforward neural network.In the feature fusion stage,the neural network first aligns the extracted features at different scales,and then extracts and fuses feature information of different ranges using separable dilated convolution with different dilation rates to finally obtain dense semantic segmentation results.The experimental results show that the generalization performance of the Conv Former is significantly improved compared with other semantic segmentation algorithms.(2)A self-training-based domain adaptation technique is proposed to migrate the knowledge learned by the deep learning model from the labeled source domain to the unlabeled target domain,so as to obtain higher performance in the target domain without using additional labeled data.Meanwhile,this thesis optimizes the method from three aspects: image input,model parameters,and learning process: firstly,cross-domain mixing of images in the source and target domains is performed to increase the image diversity without using additional annotations? secondly,semi-supervised model integration techniques are used to smooth the parameters of the deep learning model and thus stabilize the generation of pseudo-labels? then the self-training methods are improved with the locally sensitive pseudo-label quality and the dynamic weights used for balancing class learning,in domain-dependent and domain-independent perspectives,respectively.The experimental results show that the self-training domain adaptive algorithm based on local quality and dynamic class balancing proposed in this thesis has significant performance improvement compared with other domain adaptation algorithms.