Research On Hyperspectral Band Selection Method Based On Deep Neural Network

Benefit from the advantage of high spectral resolution,hyperspectral image is found widely applications in military,agricultural,and geological exploration.Particularly,as one of the research hotspots,hyperspectral image classification is the foundation of many applications.However,there are several difficulties in the hyperspectral image classification.Firstly,hyperspectral data consists of tens or even hundreds of spectral dimensions.Though it provides rich spectral information to improve the discrimination ability of ground objects,it also causes some problems,for instance,curse of dimensionality and heavy burden of store and calculate.Secondly,due to some factors such as imaging conditions,spatial resolution,and overlapping ground features,some samples with similar spectral signatures may correspond to the different class,whereas some samples from the same class may have different spectral signatures.Merely incorporating spectral information to classification greatly hinders the performance of hyperspectral image classification.In addition,because of the complex land-cover distribution in hyperspectral images,the acquisition of a large size and high-quality labeled sample set is expensive and time-consuming.Such a small sample problem easily leads to overfitting of the classification model.To deal with above problems,several novel hyperspectral band selection methods based on convolution neural network and graph convolution neural network for high-precision classification of hyperspectral images are developed in this paper.The main research contents are summarized as follows:(1)A convolutional neural network based on band-wise independent convolution and hard threshold function is proposed in this paper,which combines band selection,feature extraction and classification into an end-to-end trainable network.In this network model,a band selection layer based on band-wise independent convolution is first constructed,and it performs 1′1 convolution independently for each band.Then,a hard thresholding is used to constrain the weights correspond to unselected bands to zero,thereby achieving band selection.In addition,in order to solve the problem that the zeroed weights cannot be effectively updated in the back-propagation algorithm,two solutions are proposed by using a straight through estimator and designing a coarse-to-fine loss.compared with the gradient approximation of straight through estimator,the coarse-to-fine loss can improve interpretability.Finally,in the subsequent layers of the proposed network model,a multiscale 3D hole convolution is used to extract joint spatial-spectral features,and an auxiliary classifier is constructed to help select more discriminative bands.(2)A novel graph convolutional neural network based on additional branch and joint spatial-spectral module for hyperspectral band selection and classification is proposed in this paper.First,considering the characteristics of the hyperspectral image,a feature values and spatial locations-based graph data are established according to bands and samples.Secondly,an additional branch based on the graph data is designed,where the data are constructed by regarding spectral band as the node.The additional branch calculates the band selection mask value by integrating the band neighborhood information.This mask value is embedded in the first layer of the backbone network.Then the band selection is achieved by constraining the graph convolution layer with mask value.Finally,a subsequent network model based on joint spatial-spectral module and multi-layer feature dense fusion is designed.The band selection,feature extraction and classification are integrated into a unified optimization process.(3)A sample enhancement with local spatial constraints and non-local spectral constraints and multi-layer spatial spectral feature fusion method is developed for hyperspectral image classification.In this method,a convolutional neural network model with three branches is constructed.By merging the extracted spectral features and the dual-scale spatial features from different branches in the different layer levels,it can learn the complementary information between the detailed texture information from shallow layers and abstract semantics information from deep layers.To alleviate the small sample problem,unlabeled samples with high confidence in local spatial constraints and non-local spectral constraints are selected and pre-labeled,and then the training set is extended to learn network model parameters.Among them,local spatial constraint uses the contextual information of spatially adjacent samples,and non-local spectral constraint considers sample structure information of similar spectral information in non-local regions.