Hyperspectral Image Classification Based On Spatial-spectral Representation Learning

With the development of aerospace technology,hyperspectral images have received growing attention.As a specific kind of remote sensing images,hyperspectral images contain rich spatial and spectral information due to their continuous spectral bands.However,not all information is useful for classification tasks.Thus,how to find the important spatial regions and informative spectral bands becomes the key point.Second,in hyperspectral images,sample annotation is time-consuming and laborious.How to use limited labled samples to train a classification model is a tough task for scholars.Finally,in hyperspectral images,the distribution of samples is unbalanced and the interclass similarity is high.Some classification networks have low discriminatory power for individual categories.Thus,it becomes a challenge to make sure that each category can be classified with reasonable accuracy.In this paper,based on the above-mentioned difficulties,hyperspectral images are studied in depth,and the main results achieved are as follows.For the problem of mining important spatial regions and spectral bands in hyperspectral images,a classification method based on 3D Octave convolution and spatial-spectral attention network is proposed.This method first extracts image features using an extended 3D Octave convolutional model to mine high and low frequency spatial information of the image,while learning spectral contextual relationships.Second,inputing the extracted features into spatial and spectral attention models for highlighting important spatial locations and spectral bands in the features.Finally,an information interaction channel is established between spatial and spectral features to ensure that important spatial and spectral information is retained,while removing redundant information.Then,the fusion features are used to classify hyperspectral images.For the problems of low classification accuracy and complex network structure of individual categories of hyperspectral images,a classification method based on multi-scale spatial features and spectral attention features is proposed.The method first uses the feature learning module to extract shallow features of the image,and then feeds the extracted features into the multi-scale spatial-spectral module.This module contains the spatial Mask model and the spectral attention model.The spatial Mask model is used for obtaining multi-scale spatial features,and the spectral attention model is used for extracting spectral attention features of images.Finally,the multi-scale spatial features and spectral attention features are feed into the feature reduction module to obtain the depth semantic features of the image.Then,this feature is employed in the final classification task.Aiming at the problem of limited labeled samples in hyperspectral images,a semi-supervised classification method based on spatial-spectral graph convolution is proposed.This method first uses the SLIC algorithm to segment the original hyperspectral image into multiple superpixel blocks.In each superpixel block,every pixel point is regarded as a node to construct a spectral graph convolution model for mining the spectral features of the image.Also,due to the properties of the adjacency matrix in the graph convolution,the local spatial information of the image can be explored.Then,the spatial graph structure is built with the feature vector of each superpixel block to explore the global spatial information of the image.Finally,the feature transform model is used to convert region-level features to pixel-level features,and the pixel-level features is used to the classification tasks.