Hyperspectral Image Classification Methods Based On Joint Representations Of Structured Features Via Deep Convolutional Neural Networks

Hyperspectral remote sensing integrated spectroscopy and imaging technology realizes the simultaneous acquisition of spectral information and spatial images,which has become a significant breakthrough in the field of remote sensing.The imaging spectrometer synchronously samples the target scene with continuous narrow bands and generates the hyperspectral images(HSIs).HSIs contain rich spectral and spatial information,which can meet the detection requirements for the composition and morphology of the target object and have outstanding advantages in pixel classification and object recognition.Therefore,hyperspectral remote sensing is widely used in the fields of precision agriculture,geological survey,environmental monitoring and biotechnology.With the continuous development of imaging technology,the spectral resolution of hyperspectral images also keeps improving,which not only provides high-quality spectral information,but also puts forward new challenges for high-dimensional data processing.To sum up,developing efficient HSI intelligent processing methods to achieve refined and automated information interpretation has essential research significance and practical application value.As the critical content of hyperspectral data interpretation,HSI classification has always been the research hotspot in the field of remote sensing.Deep learning has replaced traditional methods and has become the mainstream algorithm for hyperspectral classification due to their powerful representation capabilities and flexible architectures.However,facing the multiple limitations of uncertain spectral information,complex spatial distribution,and lack of labeled samples in HSIs,extracting effective characteristics and improving the classification performance is still a challenging problem.This dissertation takes the deep convolutional neural network as the tool and focuses on the structured representation of HSIs to conduct a series of research works,which start with the cooperation and complementation of various features,optimizing the modeling mechanism and enhancing the spatial perception ability to improve the spectral-spatial representation and the classification performance.The main works and research results of the thesis include five aspects.(1)To improve the representational ability of deep convolutional networks and reduce the dependence of parameter optimization on training samples,the multi-scale alternating update clique network(MSCN)is proposed for HSI classification.Existing networks improve their data abstraction capabilities by increasing the depth and width.This way may lead to insufficient parameter optimization under limited training samples and affect the effectiveness of representation.To this end,the MSCN with the cross-loop structure is designed,which can perform cross-scale information interaction and adaptive local context extraction by skip connections between different scales,and utilize the feedback mechanism to increase the computational level and guide multi-scale features refinement.The crossloop structure enhances the feature extraction capability of the MSCN by improving the diversity and abstraction of features without deepening the existing depth of the model.In addition,parameter reuse improves model efficiency and reduces the dependence of parameter training on labeled samples.Experimental results illustrate that the MSCN can improve the discrimination of classification features,and obtain superior classification performance under limited training samples.(2)To address the issue that the limited spatial perception range of convolutional neural networks makes it difficult to capture long-range dependencies,a fully convolutional network based on class feature fusion(CFF-FCN)is proposed for HSI classification.Mainstream convolutional frameworks take the neighborhood cube of target pixels as the input and model spatial relationships through local connectivity pattern.The spatial perception range is double-limited by the input size and the convolution kernel size.The proposed CFF-FCN realizes multi-level spatial relationship learning from local to global by designing the local feature extraction block(LFEB)and the class feature fusion block(CFFB).Among them,the LFEB retains low-level detail information and high-level abstract features through loop computations to realize the complementary advantages of multi-layer information.The CFFB exploits the coarse classification mechanism to capture the global class features,and constructs the similarity measure from the class centers to each pixel to reconstruct the pixel feature and enhance its reliability.In addition,the CFF-FCN gets rid of the limitation of input size on learning spectral-spatial features through the fully convolutional framework.Experimental results show that the cooperation and complementation of local and global information can improve misclassification phenomenon and the classification performance of HSIs.(3)To solve the problem that ignoring the spectral-spatial structure prior leads to the limited discriminability of classification features,a dense convolutional network based on efficient spectral-spatial attention(DCN-ESSA)is proposed for HSI classification.Through in-depth analysis of the working principle of the existing attention mechanism,we proposed a unified attention paradigm and corresponding framework structure,and design a novel efficient spectral-spatial attention module(ESSAM)to fully utilize the inherent priors of HSIs.The DCN-ESSA combines the ESSAM with a densely connected backbone network to achieve layer-by-layer extraction and deep refinement of spectral-spatial features.The ESSAM is composed of a parameter-free spectral attention module(PFSAM)and a memoryenhanced spatial attention module(MESAM),which implements a global spectral-spatial domain with less parameter cost by optimizing feature representation,deepening information interaction,and resetting weight response.The ESSAM achieves the balance between performance and complexity while fully exploiting spectral and spatial priors,and has better interpretability and generalization capabilities.Specifically,the proposed PFSAM implements global interaction with multi-scale structured channel representation according to the characteristics of the spatial distribution in HSIs,and alleviates the insufficient channel representation.The MESAM exploits the low-dimensional properties of HSIs to capture key information with learnable memory units and enable spatial global interactions by building dependencies of learnable memory units and pixels.The experimental results indicate that applying the structural prior of HSIs to guide the network architecture design can achieve more reasonably and efficiently global optimization and further enhance the discriminative of classification features.(4)Aiming at the problem of insufficient characterization for the spatial structure of boundary samples,a multi-directional convolutional network(MDCN)is proposed for HSI classification to improve the local spatial modeling ability.The convolutional neural network places the target pixel at the center of the filter and extracts the local spatial relationship layer by layer.However,the full-window calculation for the boundary sample will introduce irrelevant features and increase the information diffusion between different categories.To this end,we proposed the side window convolution to split the full-window filter into halfwindow filters to capture the spatial relationship in different directions.On this basis,we integrated multiple side-window filter kernels into a unified convolution architecture,and further constructed the MDCN with dense connections.The MDCN can adaptively learn multi-directional spatial-spectral features and improve the representation of complex spatial structures.Experimental results show that the MDCN can improve the edge confusion phenomenon and the classification performance.(5)Aiming at the problem that the networks focus on the feature extraction of intraclass samples and ignore boundary samples,we proposed the multi-directional contrastenhanced convolutional network(MDCECN)for HSI classification.Most of the misclassifications in HSIs are concentrated in edge pixels.If the intra-class samples and edge samples are regarded as “positive” and “negative” samples,HSI classification faces the class-imbalance problem.The classification networks pay too much attention to “positive”samples and neglect “negative” samples,which is also one of the reasons for the edge confusion.To this end,the MDCECN extracts efficient spectral-spatial features with the multi-scale feature extraction network for the artificially constructed multi-directional data and achieves detail structure screening and reliable semantic enhancement through the spectral-spatial multi-directional contrast module(SSMDCM).SSMDCM exploits feature mirror transformation and the inherent transformation sensitivity of convolutional neural networks to perform vertical and horizontal contrast on multi-directional features and seek detail and invariant features,thereby driving the network to focus on reliable detail structures and complex edge pixels.Experimental results show that MDCECN can improve the classification accuracy for difficult-to-classify samples and obtain superior classification results.