Research On Hyperspectral Image Classification Based On Deep Learnin

Hyperspectral remote sensing technology combines the spectral curve of objects with ground spatial images to achieve accurate identification and attribute analysis of objects.The hyperspectral image obtained by the spectrometer not only contains rich spectral information reflecting the physical characteristics of objects,but also provides rich spatial information reflecting the spatial distribution and texture features of objects.Based on these advantages,hyperspectral remote sensing has been widely used in precise agriculture,crop monitoring,land resource utilization,urban planning and other fields.In terms of environmental protection,hyperspectral images have been used to detect changes in atmosphere,oil spills,water quality and vegetation coverage.However,the spectral characteristics of hyperspectral remote sensing are difficult to extract accurately,and the labeled samples are difficult to mark.Therefore,one of the main focuses of hyperspectral image classification tasks is to improve classification accuracy through effective feature extraction methods and small sample calculations.In recent years,deep learning methods have performed well in feature extraction of images.They can effectively solve nonlinear problems and are widely used in many image processing tasks.Based on the excellent performance of deep learning in the field of computer vision,it has been introduced into the field of hyperspectral image classification and has made great progress.This paper studies how to effectively extract and fuse spatial and spectral feature information based on three-dimensional convolutional neural networks and Transformer models to improve the classification accuracy of neural network models for hyperspectral images.The main research contents are as follows:(1)Introduce attention mechanism into convolutional neural network(CNN),propose a multi-scale feature three-dimensional convolutional neural network based on convolutional attention module(CBAM)-MSCBAMNet.The network first convolves three-dimensional image blocks into features of different sizes twice,then passes the features through CBAM modules in turn to enhance the weight of important channel information and make full use of spectral information and spatial information.Then it fuses the two-dimensional features obtained by dimensionality reduction to obtain multi-level effective space-spectrum features and improve classification accuracy.(2)Due to the attention mechanism of convolutional neural network based on local receptive fields,it is difficult to obtain long sequence context information interaction.Therefore,an improved Swin Transformer model MASwin Net based on multi-scale features and mobile window spectral self-attention mechanism is proposed to solve the problem of global and local information acquisition in hyperspectral image classification tasks.In this model,a spatial feature extraction module is used to obtain spatial information.A multi-scale spectral fusion residual module and a spectral Swin Transformer module are used to model spectral features.Finally,a fully connected layer is used to output classification results,which effectively enhances classification effect under small sample conditions.This paper conducts experiments on three representative datasets: Indian Pines,University of Pavia and Salinas,and quantitatively evaluates classification results using overall classification accuracy,average classification accuracy and Kappa coefficient.The overall classification accuracy of MSCBAMNet model on three datasets reached98.35%,98.97% and 99.14%,respectively.The overall classification accuracy of MASwin Net model on three datasets reached 98.92%,99.59% and 99.50%,respectively.The experimental results show that the two improved models proposed in this paper have achieved higher accuracy in hyperspectral image classification tasks,and MASwin Net has better classification effect than MSCBAMNet due to its combination of CNN and Transformer advantages.The experimental results also show that the two models proposed in this paper are more suitable for small sample classification,which proves the effectiveness of this method.