Hyperspectral Image Classification Based On Multiscale Feature Extraction And Lightweight Network

With the rapid development of remote sensing technology,hyperspectral images have shown great potential in the classification of ground objects.It plays an important role in agricultural planning,disaster prevention and reduction,environmental monitoring and other fields.Unlike natural images,hyperspectral images contain rich spatial structure information and a large number of spectral information bands,which makes accurate classification of ground objects possible.At the same time,there are still many unresolved issues in hyperspectral image classification through the construction of networks,such as large differences in the size of ground object categories,limited sample numbers,insufficient feature extraction,excessive model redundancy parameters,and overly complex calculations.Therefore,how to effectively extract spatial and spectral features of hyperspectral images,and design lightweight networks are key research topics in hyperspectral image classification tasks.In recent years,with the rapid development of convolutional neural networks,hyperspectral image classification methods based on deep learning have received widespread attention.On the basis of convolutional neural network,aiming at the above problems in hyperspectral image classification,this paper studies and proposes new networks and algorithms from the perspective of multiscale feature extraction and lightweight network design.The main work contents are as follows:(1)Aiming at the problems of ground objects with multiple spatial coverage sizes in hyperspectral images,and the insufficient feature extraction ability of traditional convolutional neural networks,this paper proposes a deep learning network for extracting multiscale spatial spectral features in a fine-grained manner.The core of this network is the multiscale convolution layer,which uses a multi-branch structure to achieve inter-group feature reuse,and combines depthwise 3D convolution to efficiently extract multiscale spatial spectral features from hyperspectral images.At the same time,a feature fusion module is designed to fuse information between channels using 2D pointwise convolution to achieve the joint application of high-level and low-level information,enabling the extraction of multi-scale features with stronger expression capabilities.Experimental results show that the classification performance of the proposed multiscale spatial spectral feature extraction network is superior to existing networks,and good results are obtained even in the case of fewer samples.(2)The existing methods based on convolutional neural networks usually improve classification accuracy by increasing the number or width of the network,which also brings about problems such as multiple training parameters and high computational complexity.To solve this problem,this paper proposes a lightweight multiscale deep convolutional network based on channel attention mechanism.The network uses depthwise convolutions with different convolution kernels on each branch to design a multiscale depthwise convolution and constructs a lightweight network,which fully extracting multiscale features of hyperspectral images while reducing network complexity.In addition,the adoption of a multiscale channel attention mechanism effectively captures nonlinear information across channels in feature maps,strengthening the weight of important features and improving the representation ability of different channels.Experimental results show that the proposed network not only performs well in classification accuracy,but also outperforms existing lightweight networks in terms of parameter numbers and computational costs.