Research On Hyperspectral Image Classification Method Based On Deep Semantic Segmentation Technology

Hyperspectral image classification is an important branch of remote sensing image processing research.However,as hyperspectral images themselves are characterized by high data dimension,high degree of redundancy,complex spatial-spectral structure,etc.,how to effectively improve the classification performance is still a huge challenge of hyperspectral image classification researches.In the previous studies for hyperspectral image classification based on patch center pixel,due to the overlapping of test patches and training patches during data partition,potential information leakage is caused,and the classification performance is often too optimistic.Therefore,from the perspectives of feature extraction,feature screening and information interaction,this thesis focus on a method to improve the overall classification performance by using limited spectral and spatial information without information leakage.The details are as follows:Aiming at the potential information leakage and the problem of fewer labeled pixels,a new dataset partition method is designed to avoid information leakage.First,the original image is divided into training blocks and test blocks with the same size and non-overlapping.Then,according to the total number of pixels of each type,the training blocks and test blocks are evenly divided into the training patches and test patches using a sliding window strategy,so as to ensure the objectivity of the model performance and the fair comparison between various algorithms.In addition,a variety of data augmentation strategies and their combined strategies are used to expand the number of training patches to increase the total number of samples.All the labeled pixels in the training patches obtained by this data partition method can participate in the training,which not only improves the utilization of labeled pixels,but also fits the pixel-level hyperspectral image classification method based on deep semantic segmentation technology.Given the problems of limited performance due to insufficient utilization of spectral and spatial information,a dual-attention-guided interactive multi-scale residual network(DA-IMRN)is proposed.On the one hand,a dual-attention mechanism including spatial-channel attention and spectral attention is proposed in this thesis to interactively guide the direction of feature learning in the feature extraction process,enhancing useful information and weakening useless information while realizing the interaction of spectral and spatial information.On the other hand,a multi-scale spectral/spatial residual module is proposed,which uses multi-scale residual convolution group to replace traditional convolutional layers.This module can extract spectral and spatial features corresponding to different receptive fields under limited samples to improve model performance.Experimental results on three public benchmark datasets demonstrate that attention-guided multi-scale feature learning can effectively explore spectral-spatial joint information.The overall accuracy of the proposed method is 91.26%,93.33% and 82.38%,respectively.The average accuracy is 94.22%,89.61% and 80.35%,respectively.And the Kappa coefficients are 0.885,0.923 and 0.791,respectively.All the three evaluation indices are better than the current state-of-the-art hyperspectral image classification methods without information leakage.Considering that there are few systems for hyperspectral image classification,the above-mentioned dataset partition method and image classification network are packaged and a specialized system for hyperspectral image classification is designed based on the C# programming language.It includes five modules: image reading,image preprocessing,image classification,performance evaluation,and data storage/query.It has preliminarily explored the application of research results in practical production.This thesis has 35 figures,15 tables and 88 references.