Research On Coastal Land Cover Classification With High-resolution Remote Sensing Images Based On Deep Learning

As an important survey content,the coastal land cover classifications directly reflect the results of the joint effects of the sea-land climate and human activities in this special zone,and have practical application value in sustainable development research such as land planning,resource management,and ecological environment protection.The rapid development of earth observation and remote sensing technology has made it possible to obtain massive amounts of high-resolution remote sensing images,and has also provided opportunities and challenges for large-scale and fine-grained land cover classification research in coastal areas.When it is difficult for traditional machine learning algorithms to manually extract effective features to map the correct land cover category for each pixel,the deep learning technology has shown a wide range of applications in the field of remote sensing image processing by virtue of its strong ability to describe nonlinear features,and is gradually applied to the coastal land cover classification tasks.However,the classification models based on deep learning still have some limitations in the network structure,and their effectiveness always relies on a large amount of label data that needs to consume huge time and labor costs.Therefore,this thesis first aims to establish a coastal land cover classification system and classification datasets on the basis of in-depth investigation and understanding of the characteristics of land cover categories in high-resolution remote sensing images.On the other hand,this thesis hopes to reasonably introduce new deep learning methods to achieve data-efficient coastal land cover classification results for the different supervised,semi-supervised or unsupervised data scenarios,and ultimately provide a high-precision geographic reference for the follow-up key research area.The main work and contributions of this thesis are as follows:(1)This thesis systematically expounds on the scientific value and development trend of using high-resolution remote sensing images to carry out the coastal land cover classification,and analyzes the influence of various factors.According to the different scenarios that the remote sensing data may present,the research status and existing bottlenecks of deep learning methods for land cover classification under supervised,semi-supervised and unsupervised conditions are highlighted respectively.(2)This thesis comprehensively expounds on the feasibility and superiority of deep learning technology in the coastal land cover classification with high-resolution remote sensing images,in which the key technologies involved in the convolutional neural network structure,model training process and classic classification algorithms are sorted out and summarized in detail that provides a solid theoretical and technical support for the construction of the coastal land cover classification model under different data scenarios.(3)In view of the data demand of coastal land cover classification research,this thesis has developed several high-resolution benchmark datasets for land cover classification in line with the basic realities of coastal areas.On the basis of a series of pre-data preparation work such as the selection of the research areas,the acquisition of remote sensing images,and data preprocessing,a land cover classification system suitable for high-resolution remote sensing images in this special area was constructed under the actual research purpose,and two sets of benchmark datasets for land cover classification were finally formed combining with the annotated refined ground truth.Thus,this work provides and supplements the data foundation in the research field of coastal land cover classification based on deep learning technology.(4)Aiming at the local limitation in the context and the difficulty of recovering the spatial details of classification models in supervised remote sensing data scenes,this thesis proposes a coastal land cover classification method based on multi-attention coding.First,a position-channel attention joint module is introduced for the high-level semantic feature from the encoding path,in which the location-relative attention module and the channel-relative attention module encode the global context in the spatial and channel dimensions to enhance the feature representation of coastal land cover categories with dependencies respectively.Secondly,a lightweight global feature attention module is proposed and applied to multiple feature scales to form a complete decoding path,which in turn utilizes the global context generated by high-level semantic features to guide the fusion of spatial details contained in low-level features.Extensive experiments show that the proposed method successfully integrates the global context information from high-level semantic features and the local detail information from low-level features,and has achieved excellent performances in the overall performance and classifying the coastal land cover categories with diverse and complex features.(5)Aiming at the model performance degradation that may be caused by insufficient labeled samples in the semi-supervised remote sensing data scenes,this thesis proposes a semi-supervised coastal land cover classification method based on multiple pseudo-labels.First,the consistency regularization is introduced,and the perturbations with different forms are respectively imposed on the multi-scale features with more obvious decision boundaries to enforce the consistency constraints of predictions on unlabeled training samples,in which the one-hot classification results from the auxiliary classification network assume the role of pseudo-labels of the main classification network.Secondly,a self-training scheme with a simple mechanism is proposed to automatically select and prioritize more reliable and easier pixels for unlabeled data and generate pseudo-labels with stability,so as to further improve the coastal land cover classification results under the semi-supervised data scenarios.Compared with other methods,it is proved that the proposed method can more effectively locate the decision boundary to achieve data-efficient classification effects,and finally generate the prediction results with higher confidence in multi-scale coastal land cover categories and boundary areas.(6)Aiming at the problem that the classification models may be difficult to train due to the complete absence of labeled samples in the unsupervised remote sensing data scene,this thesis proposes an unsupervised coastal land cover classification method based on class-aware domain adaptation.First,according to the generative adversarial network technology,the global adversarial domain adaptation module and the local adversarial domain adaptation module are introduced to perform edge distribution and joint distribution alignments in the output space and feature space,respectively.Secondly,the entropy minimization module is proposed to directly generate high-confident predictions for unlabeled training data in the end-to-end training process without introducing additional hyperparametric constraints.On the other hand,a comprehensive loss re-weighting scheme is proposed to solve the inherent problems of high-resolution remote sensing image data in coastal areas.A series of comparison and ablation experiments show that the proposed method can successfully learn the category features after data domain conversion,and finally generate accurate coastal land cover classification results,especially for some land cover categories that present minority and difficult characteristics.