Hyperspectral Image Classification Based On Deep Learning And Attention Mechanism

Hyperspectral image classification(HSI)is one of the most important research projects in remote sensing image processing tasks.However,due to the high dimensionality and redundancy coming from hyperspectral image itself and the imbalance of classes in HSI dataset,how to improve the classification performance of network is still a huge challenge.In recent years,deep learning-based methods have been widely used in hyperspectral image classification tasks,but these methods do not perform well in classification of some classes.Therefore,how to make full use of the spatial and spectral information,how to ensure efficient network training,and how to reduce the influence of noise from HSI data itself and the classification process are three key issues for HSI classification tasks.This paper conducts in-depth research on the above issues,mainly including the following aspects:(1)Aiming at the problem of how to make full use of the spatial and spectral information and solve the imbalance of classes,an end-to-end 3D-CNN-based 3D-Res Ne Xt network is designed.This network also uses feature fusion and label smoothing strategies.On the one hand,residual connection and split-transform-merge strategy can alleviate the declining-accuracy phenomenon and greatly reduce the number of parameters.Only by adjusting the number of packets of the packet convolution instead of the network depth,more distinctive features can be extracted and the classification performance of the network can be improved.On the other hand,in order to improve the classification accuracies of classes with small numbers of samples,we enrich the input of the 3D-Res Ne Xt spectralspatial feature learning network by additional spectral feature learning,and finally use a loss function modified by label smoothing strategy to solve the imbalance of classes.(2)Aiming at the problem of how to ensure efficient network training,a 3D-CNN-based residual group channel and space attention network(Residual Group Channel and Space Attention Network,RGCSA)is designed.On the one hand,the proposed attention mechanism with residual connection can improve the training efficiency of network by optimizing the features of channel dimension and spatial dimension during the whole training process.On the other hand,the proposed residual group channel-wise attention module can reduce the possibility of losing useful information,and the novel residual spatial-wise attention module can extract context information to enhance spatial features.Our proposed RGCSA network only needs few training samples to achieve higher classification accuracies than previous 3D-CNN based networks.(3)Aiming at the problem of how to reduce the influence of noise from HSI data itself and the classification process,a 3D-CNN based Two-Stage Attention Network(TSAN)for HSI classification was proposed.On the one hand,the spectral-wise attention module in the first stage can optimize the whole spectrum by shielding useless spectrum bands and reducing the noise in the spectrum.On the other hand,more discriminative spectral-spatial features are extracted and sent to the subsequent layers by channel-wise attention mechanism combined with soft thresholding in the second stage.In addition,by introducing non-local block to learn global spatial features,multi-scale network can be used to combine the local space and the global space.Compared with the previous 3D-CNN-based networks with attention mechanism,the proposed TSAN can indeed reduce the influence of noise from HSI data itself and the classification process,and only needs less training time to achieve higher classification accuracy.