Research On Remote Sensing Image Classification Based On Deep Networks

The current Earth observation system focuses on the collaborative use of the complementarity between multiple imaging sensors to achieve all-sided observation of the ground scene with high spatial resolution and high spectral resolution.Increasingly mature remote sensing imaging technology has brought massive amounts of optical remote sensing images.How to accurately and quickly interpret these images to serve the applications,such as environmental monitoring,geological exploration,and urban planning,has become the research hotspot in the remote sensing community.Based on the characteristics of the high spatial resolution and high spectral resolution of the Earth observation system,this dissertation conducts the study of scenelevel and pixel-level remote sensing image classification,respectively.Scene-level remote sensing image classification refers to the high-resolution remote sensing image classification,and the pixel-level remote sensing image classification refers to the hyperspectral image classification.Due to its powerful representation capabilities,deep learning-based methods have shown superior performance.However,these methods still face following problems: 1)Unsupervised feature encoding methods are difficult to effectively explore the semantic labels to promote the aggregation of multi-layer convolutional features;2)Existing feature aggregation methods usually ignore the interference information between multi-layer convolutional features;3)Existing methods usually exploit convolution operations to build deep networks,which are sensitive to the spatial rotation of hyperpectral images;4)Existing convolutional network-based methods heavily rely on the assumption that neighboring pixels have the same category.These methods tend to overfit spatial distributions of land-covers and have poor generalization performance.In response to above-mentioned problems,this dissertation carries out the research on the remote sensing image classification from four aspects.The main research contents and contributions are introduced as follows:(1)Feature aggregation network for remote sensing scene classification.Since unsupervised feature encoding methods are difficult to effectively explore the semantic labels to promote the fusion of complementary information among multi-layer convolutional features,this dissertation proposes to integrate feature learning,feature aggregation and classifier into a unified end-to-end framework.A convolutional feature encoding module is designed to adaptively aggregate multi-layer convolutional features,which can be embedded in deep networks for joint optimization.A progressive aggregation strategy is proposed to extract discriminative scene representation.Compared with existing methods,overall accuracy of the proposed method on the WHU-RS19 dataset is improved by 1.7%.(2)Gated bidirectional network for remote sensing scene classification.Due to that various convolutional features may contain interference(mutual exclusion or redundancy)information,this dissertation integrates the aggregation of multi-layer convolutional features and the elimination of interference information into a deep network.A bidirectional connection is exploited to aggregate multi-layer convolutional features into the semantic-assist feature and the appearance-assist feature at the topdown and bottom-up directions,respectively.A gating function is proposed to suppress the interference information among various convolutional features.Compared with existing methods,overall accuracy of the proposed method on the WHU-RS19 dataset is improved by 2.3%.(3)Hyperspectral image classification based on rotation-invariant attention network.Since there are redundant spectral bands in hyperspectral images,this dissertation proposes a central spectral attention module to suppress the redundant spectral bands in a task-driven manner and improve the discriminability of subsequent features.A rectified spatial attention module is proposed to replace the traditional convolution to extract rotation-invariant spectral-spatial features.Finally,these two attention modules are utilized to build the rotation-invariant attention network to classify hyperspectral images.Compared with existing methods,overall accuracy of the proposed method on the Pavia University dataset is improved by 4.4%.(4)Hyperspectral image classification based on fully convolutional segmentation network.Due to that the hyperspectral image usually contains complicated spatial distributions of land-covers,this dissertation proposes a fine labeling method to label the categories of all pixels in hyperspectral image to obtain detailed spatial distributions of land-covers.A hyperspectral image generation method is designed to improve the diversity of spatial distributions in training set.When the spatial distributions of landcovers on the Pavia University dataset change,overall accuracy of the proposed method is improved by 8.8% compared with existing methods.