Research On Classification Of Hyperspectral Remote Sensing Images With Multiple Kernel Learning

Nowadays,hyperspectral remote sensing has become an important means of Earth observation and even space exploration.Hyperspectral image,as a complete entity of image and spectrum,can not only capture the detailed spectral information,but also provide the spatial distribution of the scene of observation,affording a great potential for more accurate identification of the landcover types of interest.Therefore,as the core means of hyperspectral remote sensing image analysis,spatial-spectral classification has always received extensive attention and research,and has been widely used in related fields,including national land resources survey,precision agriculture and forestry,urban land use planning,environmental protection and so on.Recent years,with the miniaturization of hyperspectral sensors and the popularity of unmanned aerial vehicles,the spatial resolution of hyperspectral images has been continuously improved and the repeated observation of the same area has become more convenience.At present,the classification of hyperspectral remote sensing images faces three main scientific problems: the spatial-spectral multi-feature heterogeneity of the hyperspectral images,the complexity of the spatial distribution of landcover types under the condition of high spatial resolution,and the spectral uncertainty caused by cross-temporal hyperspectral imaging under different observation conditions.Aiming at the above scientific problems,based on the multiple kernel learning theory,this paper carried out research work from three aspects: pixel-level spatial-spectral classification of hyperspectral image,object-level spatial-spectral classification of hyperspectral image,and transfer for classification of cross-temporal hyperspectral images.The main work includes the following contents::(1)Aiming at the problem of spatial-spectral multi-feature heterogeneity of hyperspectral images in pixel-level classification,take the nonlinear interaction between features into consideration,a method of heterogeneous information fusion based on spatial-spectral multi-feature nonlinear multi-kernel learning is proposed.Firstly,multi-scale and multi-structure elements are used to detect the geometric structure of hyperspectral images,and multi-morphological profile feature extraction of hyperspectral images is realized;Furthermore,each morphological profile is used to construct a linear base kernel,and a nonlinear combinatorial learning algorithm of these base kernels is designed.The similarity and discrimination information of different scales and different structures in the feature space is effectively mined through the nonlinear coupling of base kernels,and the fusion of spatial-spectral features with different scales and structures is realized,so as to improve the measurement ability of similarity and discrimination between samples,and greatly increase the classification accuracy.(2)In order to solve the problem of interpretation of classification and discrimination ability of spatial-spectral multi-feature in pixel-level hyperspectral image classification,a heterogeneous feature interpretation method based on class-specific sparse multi-kernel learning is studied.Because the spatial-spectral features with different degrees and attributes extracted from the original hyperspectral image have different effects on classification task,therefore,how to recognize and analysis the discriminate ability of different spatial-spectral features is a critical issue.First of all,the multi-degree and multiple attribute filters are used to extract the multi-level spatial-spectral features of hyperspectral images,and corresponding Gaussian base kernels are generated.Next,a class-specific sparse multiple kernel learning method under the constraint of Group LASSO is proposed,which can recognize the contribution degree of different spatial-spectral features to the final classification performance of a given pair of catergries.Thus,feature selection and feature interpretation are realized.At the same time,redundant features can be eliminated while ensuring classification performance.(3)For the complexity of the spatial distribution of landcover types under the condition of high spatial resolution in object-level hyperspectral image clssification,multi-attribute superpixel model and multi-attribute superpixel-tensor model is put forward to realize the adaptive and fine extraction of spatial-spectral features of high spatial resolution hyperspectral images.Under the condition of high spatial resolution,the boundaries of different categories are clearer,which make it possible to realize fine classification of land cover types using spatial features.However,the more refined spatial morphology of these categories makes the traditional window-based feature extraction method no longer applicable.Meanwhile,more abundant texture will cause serious over-segmentation.In this paper,superpixel segmentation is performed on the extended multi-attribute profiles of the hyperspectral image to reduce intra-class differences,weaken the influence of false boundaries on segmentation.Secondly,when extracting features between multi-attribute superpixels,consistency criteria are introduced to merge similar superpixels,ensuring the correctness the many-to-one mapping relationship between the merged superpixels and the ground object category,so that the classification accuracy can be increased.Based on the multi-attribute superpixels,multi-attribute superpixel-tensor model is constructed,that is,the tensor representation of each multi-attribute superpixel is carried out,followed by the extraction of features with unified dimension,which not only maintains the adaptability and uniformity of the spatial structure brought by the superpixel,but also considers the coupling relationship between each dimension of the superpixel,conduciving to improve classification accuracy.(4)Aiming at the problem of spectral uncertainty caused by the difference of illumination,atmosphere or hyperspectral sensors under different observation conditions,a domain adaptive classification model,i.e.,spatial-spectral multiple geodesic flow kernel learning,is proposed.When hyperspectral images are obtained from different temporals,the uncertainty of illumination,atmospheric conditions or different sensors leads to a large spectral difference in the cross-temporal hyperspectral images,meanwhile,labeling the hyperspectral data is difficult.This paper studies a domain adaptation method for classification of cross-temporal hyperspectral data cover the same scene,multiple geodesic flows are constructed based on each pair of spatial-spectral features from source and target domain,followed by the construction of corresponding Gaussian geodesic flow kernels.By embedding the multiple geodesic flow kernels into multi-kernel learning framework,the spatial features are integrated into the domain adaptive model for classification.Therefore,the low-dimensional intrinsic structure of hyperspectral data is mined and the domain-invariant multi-features are obtained,so as to improve the accuracy of classification after domain adaptation.