Research On Hyperspectral Image Classification Based On 3D Convolutional Neural Networks And Attention Mechanism

Hyperspectral image（HSI）is a kind of high-dimensional and special remote sensing image,which is a comprehensive information carrier integrating spatial and spectral information.Compared with ordinary remote sensing images,HSIs contain richer information,which provide convenience for a series of subsequent computer vision tasks.HSI classification based on pixel spatial and spectral information is the most basic and meaningful task.Recently convolutional neural network（CNN）technology has advanced by leaps and bounds,and more and more methods for HSI classification based on CNN have emerged and achieved certain success.However,the current HSI classification still faces many problems,including less training data sets,redundant information,and prone to Hughes phenomenon,that is,with the continuous increase of spectral dimensions,the classification accuracy improves at first and then drops sharply.This paper combines with the CNN technology to carry out the following research on account of these problems.（1）A multi-scale densely connected 3D CNN for HSI classification is proposed in this paper.Three branches with different sizes of 3D convolution kernels are designed in order to describe the hierarchical spatial-spectral characteristics.So as to alleviate the disappearance of the gradient,a dense connection structure is designed in each branch,and the input of each layer contains all the feature maps generated by the previous layers.The proposed network contains only five convolutional layers,so the network structure is short and concise.Experimental results on two public datasets show that this method has a significant improvement over the existing methods and requires less calculation and implementation time,and also performed well in small sample types.（2）Using the conventional 3D CNNs to extract the spatial-spectral feature for HSIs results in too many parameters as HSIs have plenty of spatial redundancy.To address this issue,this paper first designed multi-scale convolution to extract low-level features of different scales of HSIs.What is more,octave 3D convolution is used in the multi-scale module to replace ordinary 3D convolution to reduce spatial redundancy and expand the receptive field.In order to further dig out the distinguishing features,the channel and spatial attention modules are used to optimize the feature map and improve the classification performance.Experimental results on four publicly available datasets show that the proposed method can effectively remove information redundancy and obtain more discriminative features in HSIs,thus improving classification accuracy.