Research On Classification Method Of Hyperspectral Remote Sensing Image Based On Deep Learning

Hyperspectral image is a kind of special image because of the information of hundreds of continuous spectral bands and rich spatial information.However,the redundant information brought by the high-dimensional and continuous spectral bands and the interference pixels in the spatial neighborhood will affect the classification accuracy.In addition,the lack of labeled samples makes deep network unable to train well,which limits the development of hyperspectral image classification.In this paper,the above problems faced by hyperspectral image classification are explored,and the main work and contents are as follows:（1）Faced with the effect of spectral redundancy information and neighborhood spatial interference pixels on classification performance,a supervised hyperspectral image classification method based on multi-attention and multi-level deep feature fusion is proposed.This method uses the band attention module when extracting spectral features to alleviate the influence of redundant spectral bands;when extracting spatial features,it uses the spatial attention module to suppress the influence of interfering pixels in the spatial neighborhood;in the joint spectral-spatial feature extraction,the spatial attention module is used again to suppress the interference pixel information introduced by the spectral feature extraction during the spatial-spectral feature fusion process.In addition,in order to reuse the features extracted from different network layers,a multi-layer feature fusion mechanism is designed for space-spectral feature fusion.The experimental results on three public datasets prove that the model has high classification performance.（2）In order to solve the problem of insufficient hyperspectral image label samples in the deep learning training process,a semi-supervised hyperspectral image classification method with false label suppression and deep dense feature fusion is proposed.First,the semisupervised learning strategy is used to make full use of unlabeled samples,and the proposed scoring mechanism is used to select unlabeled samples with higher confidence for pre-labeling.Then use pre-labeled samples and labeled samples to train the network.Secondly,a robust loss function is introduced to suppress the influence of incorrect labels on classification performance.Finally,the deep dense network is used to fuse the features to realize the multiplexing of different levels of features and alleviate the phenomenon of over-fitting in the training of the deep network.Experimental results on two public datasets prove the robustness and effectiveness of the model under limited labeled samples.