Researches On Joint Classification Methods For Multi/Hyperspectral Image And LiDAR Data

With the improving demand for classification and recognition of ground covers,new requirements are put forward for remote sensing scene interpretation: higher spatial two-dimensional(2D)interpretation accuracy,and spatial three-dimensional(3D)interpretation.The joint use of multi-source data and new technology of remote sensing are effective means to meet the escalating demands.Multi/hyperspectral imaging and monochromatic Airborne Light Detection And Ranging(LiDAR)are two main sensors for detecting information about ground covers.Multi/hyperspectral images provide 2D spatial information and rich spectral information of remote sensing scenes.However,due to the limitation of imaging principle,the spatial 3D distribution information is lost.Airborne monochromatic LiDAR(hereinafter referred to as LiDAR),as an active remote sensor,aquires spatial 3D information of remote sensing scenes,which is complementary to multi/hyperspectral image.In recent years,with the continuous development of stereoscopic spectral imaging technology,integrated acquisition of spatial 3D and spectral dimensional information of remote sensing scene has been achieved.Triple-wavelength LiDAR simultaneously acquires spatial 3D information and three-band laser pulse echo intensity information.That provides data support for 3D interpretation of remote sensing scene.Multi/hyperspectral images and LiDAR data(or three-band LiDAR data)provide more discrimination information for remote sensing scenes,but also bring new challenges for information extraction and classification.How to extract discriminative features and design classifier for multi/hyperspectral images and LiDAR data? How to design a point cloud feature extraction and classification model based on 3D space-intensity integrated information of the new three-band LiDAR data? This dissertation deeply analyzes the complementary characteristics of spatial and spectral(or intensity)dimensions.Aiming to make full use of the spatial 3D information and spectral(or intensity)information to improve the classification accuracy of ground covers,this dissertation carries out research from multi/hyperspectral images and LiDAR data joint classification,3D space-intensity information joint classification for three-band LiDAR point cloud data.The research achievements are mainly reflected in:(1)Aiming at the problems of discriminative feature extraction under strong heterogeneity of image-spectrum-elevation information from multi/hyperspectral images and LiDAR data,this dissertation proposes a feature extraction method for multi/hyperspectral images and LiDAR data.The proposed method is based on discriminative graph fusion.First of all,according to the multi-scale characteristics of ground covers in remote sensing scene,multi-scale similarity graphs are constructed in each attribute feature space to reflect the distribution within the same class.In order to increase the distribution distance between classes and enhance the separability of ground covers,non-similar graphs are further constructed using each attribute feature.Then various attribute graphs are fused into one graph.Graph fusion eliminates the false structures existing in the single-attribute feature space.The fused graph only retains the connection relationship reflecting the essential class information.Finally,for classification of ground covers,the feature projection matrix is solved by taking the maximum class separability as the optimization criterion.Then the complementary information in multi-attribute features is mapped into low-dimensional features.The mapped features provides more reliable discrimination information for ground cover classification.Compared with mainstream feature fusion methods,the proposed feature extraction method can achieve better classification performance.(2)Aiming at the multi-scale nonlinear distribution of ground covers in image-spectrum-elevation fusion feature space,a feature-scale double-layer discriminative multiple kernel learning framework is adopted to classify the ground covers.First,multi-scale basic kernels are constructed on each feature to learn the multi-scale information of ground covers.Then the multiple kernel learning on features is carried out to combine the complementary information of different features.To solve the model of multiple kernel learning,a discriminative multiple kernel learning method is further proposed.It uses the sample distribution information in kernel feature space to optimize the combination weight coefficients of basic kernels.The learned multiple kernel learning model enhances the ability of classification for ground covers.Compared with representative multiple kernel learning methods in the field of remote sensing,the proposed method improves the classification accuracy.And it can analyze the importance of the basic kernels for classification.(3)Aiming at the problem of 3D space-intensity information integrated representation and complementary information fusion for three-band point cloud,a feature extraction method based on tensor manifold discriminant embedding is proposed.Firstly,a tensor model is proposed to represent the 3D spatial information and three-band echo intensity information for three-band point cloud integratedly.The proposed tensor model retains the coupling relationship between various information dimensions.On this basis,the information of different dimensions is fused to extract the geometry-intensity fusion features.Compared with the original three-band point cloud,the extracted features improve the classification accuracy by more than 10%;Compared with the current representative feature extraction methods based on vector and tensor model,the features extracted by the proposed method have stronger classification capability.(4)Aiming at the insufficient utilization of high-order information of ground covers for the three-band LiDAR data multi-dimensional feature joint classification,a multi-attribute smooth graph convolutional network is proposed for three-band point cloud classification.Traditional feature fusion methods cannot effectively extract the invariant essential features from nonlinear data.This dissertation combines deep network structure for extracting high-order invariant features and graph model for non-linear information representation to further mine the discriminating features from three-band point cloud.Firstly,three-band point cloud is segmented into a series of superpoints.These superpoints are used as subsequent classification objects.Then,the graph model with each attribute information is established by taking each attribute feature as measurement criterion.And the multi-attribute graphs are combined as the input of smooth graph convolutional network.Finally,according to the spatial clustering characteristics in remote sensing scene,multiscale heat operator is introduced to improve the graph convolution.The introduction of heat operator improves the ability to recognize the class of points in point cloud.Experimental results on real three-band point cloud show that,compared with current representative graph convolution network methods,the proposed method improves the numerical 3D classification accuracy.