Deep Learning-based Methods For Hyperspectral Band Selection

The emergence of hyperspectral remote sensing technology is one of the most iconic achievements in the history of remote sensing development.Hyperspectral images(HSIs)realize the integration of spatial and spectral information,which makes fine ground object recognition possible.However,the excessive number of spectral bands contain some noisy and redundant ones,which have an adverse effect on the classification performance,the consumption of computing resources,transmission,and storage of HSIs.Therefore,it is necessary to reduce the dimensionality of HSIs before HSIs are processed and analyzed.Due to the great advantages of deep neural networks in nonlinear representation,some studies have applied deep neural networks to the field of hyperspectral band selection.However,it is difficult to directly calculate gradient of the objective function,and the number of labeled samples is very limited.In addition,the existing methods lack efficient heuristic search strategy for the band search space with the exponential increase of the number of bands.Also,existing methods evaluate bands by sampling using the classification network,which limits the performance and efficiency of band selection.Finally,the existing paradigm of image band selection needs to train a single model for each dataset,which undoubtedly ignores the inherent similarity between different hyperspectral image band selection task instances,resulting in a huge waste of computing.In this study,ternary weight convolutional neural network,deep reinforcement learning,meta-learning,and graph neural network are used to solve the application problems of deep learning in the field of hyperspectral band selection from the perspectives of end-to-end feature extraction and classification,unlabeled sample utilization,high-performance band search and efficient feature evaluation,and general band selection model.The main contributions of this study are listed as follows:(1)A semi-supervised hyperspectral image band selection method based on ternary weight convolution neural network(TWCNN)is proposed.Aiming at the end-to-end application of deep learning method in the field of band selection of HSIs,this method constructs a ternary convolution neural network with a weight of-1,0 or 1 to realize the selection of band.In addition,in order to alleviate the problem of small size samples in hyperspectral image processing,a semi-supervised feature extraction module is constructed,which realizes the use of a large number of unlabeled samples by constraining the compactness in the feature of the network,which also improves the discriminant ability of the deep features.Furthermore,the algorithm integrates band selection,feature extraction,and classification into a unified optimization process,and realizes end-to-end band selection and classification.(2)A band selection method for hyperspectral images based on deep reinforcement learning(RLBS)is proposed.This method formulates the hyperspectral band selection problem as reinforcement learning to achieve efficient and accurate band selection from the point of view of band search and band evaluation.In this algorithm,the band search process is modeled as a Markov decision process.After that,for the evaluation of band subset,a simple random band sampling strategy is designed to train the neural network,so that the network can be evaluated effectively without any fine-tuning.On this basis,a reinforcement learning agent is designed,which searches in the state space and expects to maximize the cumulative reward to select the best band subset,which is based on the band selection of reinforcement learning.(3)On the basis of the above two works,a hyperspectral image band selection method based on meta-learning is proposed.In view of the defect that the current method needs to train a model separately for each data set,this method models the band selection file as a metalearning problem,explores the inherent similarity between instances,and realizes the training and testing across datasets.On the basis of meta-learning modeling,this method uses graphs to represent fusion band relationships,further carries out graph embedding based on graph convolution neural network to extract cross-spectral information,and transforms band selection problem into graph node selection problem.Finally,a reinforcement learning method is used to train the model.Experimental results on the real datasets shows good performance of the proposed method.