Research On Hyperspectral Image Classification Method Under Different Number Of Labeled Samples

With the development of hardware and software such as hyperspectral imaging technology and deep learning algorithms,hyperspectral remote sensing image classification has been widely used in precision agriculture,environmental monitoring,military reconnaissance and other fields.In the field of hyperspectral image classification,according to the number of labeled samples,the classification situation can be divided into three types: sufficient-labeledsample,insufficient-labeled-sample,and no-labeled-sample.With different number of labeled samples,the main problems faced by hyperspectral image classification under these three situations are different: irregular shapes of objects leading to over-smoothed spatial-spectral features at edges;spectral variability resulting in a small number of samples unable to represent global feature distribution;differences in imaging conditions causing inconsistent feature distributions between the unlabeled image to be classified and the labeled image.To address these problems,this thesis takes "different number of labeled samples" as the main line of research,and studies how to extract accurate,discriminative,robust spatial-spectral features for hyperspectral image classification under sufficient-labeled-sample situation,insufficientlabeled-sample situation and no-labeled-sample situation respectively.The main research contents are as follows:Firstly,under sufficient-labeled-sample situation,the superpixel guided deformable convolution network for hyperspectral image classification is proposed to address the problem of irregular shapes of objects leading to over-smoothed spatial-spectral features at edges.The designed superpixel guided deformable convolution(SGD-Conv)can change the shape of the convolution kernel according to actual ground object shapes,so as to extract more accurate spatial-spectral features and make ground objects in the classification result have accurate,clear and high-fitting spatial texture structures.In order to generate highly homogeneous superpixel regions,a superpixel region fusion filter is designed,which can accurately segment texture structures of objects of different sizes and improve accuracy of SGD-Conv offset.Additionally,in order to solve misclassification problems at both pixel-level and region-level in the classification result,the superpixel joint bilateral filter and double-window joint bilateral filter are designed to correct the misclassification problem with the help of the homogeneity of the superpixel region and the expanded correction window,so that the texture of the ground object can be more closely matched to reality.The experimental results show that,in the case of sufficient-labeled-sample,this algorithm not only guarantees high accuracy classification,but also accurately classifies the texture structure of the ground objects.Secondly,under insufficient-labeled-sample situation,the multiscale short and long range graph convolutional network for hyperspectral image classification is proposed to address the problem of spectral variability resulting in a small number of samples unable to represent global feature distribution.The designed long-short graph convolution uses two parallel branches to independently perform short-range and long-range graph convolution,making the ground object features more consistent and accurate at the global range.In order to solve the problem that a single-scale graph structure do not adequately express the spatial structure of objects,a multi-scale superpixel generation sub-network is designed,which extracts spatial-spectral features of different scales to ensure full utilization of texture structures at different scales.Meanwhile,a new approach is proposed to automatically determine the number of superpixel segmentation regions based on inherent properties of hyperspectral images.Furthermore,a feature transformation sub-network is proposed to reduce the computational complexity of the graph convolution network.The experimental results show that this algorithm is able to extract more accurate and consistent global features in the case of insufficient-labeled-sample,thereby improving the classification accuracy.Finally,under no-labeled-sample situation,multilevel feature alignment based on spatial attention deformable convolution for cross-scene hyperspectral image classification is proposed to address the problem of differences in imaging conditions causing inconsistent feature distributions between the unlabeled image to be classified and the labeled image.The newly designed multi-level feature alignment method achieves feature alignment of source domain(SD)and target domain(TD)using three levels: TD sample to SD sample,TD sample to SD cluster-center,TD cluster-center to SD cluster-center.In addition,in order to extract accurate spectral features for each category,the spatial attention deformable convolution is proposed,which can extract accurate and pure spatial-spectral features for each category in the subsequent feature alignment process,avoiding unsatisfactory feature alignment due to inaccurate features of various categories.Experimental results show that this algorithm is able to achieve the identical feature distribution between the target domain and source domain,and achieve high performance classification of unlabeled target domain.