Research On Spatial Spectrum Synergy Multi-Scale Hyperspectral Remote Sensing Image Classification

Due to the advantages of high spectral resolution,hyperspectral images provide more detailed ground feature information for geoscience analysis,and are widely used in precision agriculture,pest monitoring,mineral exploration,and other fields.Hyperspectral image classification is one of the basic research directions of hyperspectral remote sensing,and is the prerequisite for many geoscientific analyses of hyperspectral remote sensing.Therefore,the study of hyperspectral image classification has very important research significance.Hyperspectral images have the advantages of high spectral dimensions,redundant information,and other issues,resulting in low classification accuracy of existing algorithms.Compared with traditional methods,deep learning based hyperspectral image classification methods can achieve higher classification accuracy,but there are also problems such as insufficient feature extraction,ignoring multiscale features of hyperspectral images,and underutilization of spectral information.Therefore,the following research was carried out in this article:(1)A 3D Pyramid Residual Network(3DPRes Net)for hyperspectral image classification is proposed due to issues such as gradient vanishing and excessive model computation caused by increasing network depth.This network introduces a residual structure deepening network based on the 3D Convolutional Neural Network(CNN)to enhance feature extraction capabilities.At the same time,the pyramid feature graph growth mode is used to control the slow increase in the number of feature maps,effectively controlling the computational complexity of the model.The overall classification accuracy of the proposed model has improved by 10.7% and11.31% respectively compared to SVM on the Pavia University(PU)and Salinas(SA)datasets,and the average classification accuracy has improved by 11.39% and 12.64%,greatly improving the effectiveness of ground object classification.(2)A multi-scale convolutional neural network(2-3D NL-CNN)based on Non local attention mechanism and Inception module is proposed to address the issues of incomplete single scale feature extraction and insufficient utilization of spectral information.The network adopts a Non local attention mechanism to capture the dependency relationship between spectral and spatial dimensions,and extract more comprehensive spatial spectral features;The Inception module provides the network with multi-scale spatial feature extraction capabilities,enabling a more comprehensive extraction of spatial information.By utilizing spatial and spectral features for classification,the overall classification accuracy of the model on PU and SA datasets reached 99.98% and 99.65%,respectively,with an average classification accuracy of 99.76% and 99.42%,achieving better classification performance compared to existing deep learning models.(3)In response to the problem that current deep learning based hyperspectral image classification methods still mainly rely on spatial features for classification,and the spectral features of pixels are not fully utilized,a hyperspectral image classification network(CNN-SPGRU)based on CNN and Recurrent Neural Network(RNN)spatial spectrum collaboration is proposed.The network adopts a dual branch structure,with the CNN branch using Principal Component Analysis(PCA)to reduce the dimensionality of hyperspectral images and extract spatial features.The RNN branch uses shorter Parallel Gate Recurrent Unit(SPGRU)to extract spectral features.The spatial and spectral features are concatenated and input into the fully connected layer to learn the linear combination of the two,thereby achieving hyperspectral image classification through spatial spectral collaboration.Compared with other models,the overall classification accuracy of the model improved by 0.21%～0.6%and 0.44%～0.81% on PU and SA datasets,and the average classification accuracy improved by 0.11%～1.34% and 0.08%～0.29%.(4)Comparative analysis was conducted on the models proposed in this article to evaluate the performance and applicability of the models.In terms of model stability,the 3DPRes Net and 2-3D NL-CNN models have better stability than the CNNSPGRU model,and exhibit stronger robustness to "homospectral foreign matter" and "homospectral heteromorphism" phenomena in hyperspectral images;In terms of training time and accuracy,the computational complexity of the CNN-SPGRU model is much smaller than that of the other two models;From the perspective of model small sample learning effects,the feature extraction capabilities of 3DPRes Net and 2-3 D NL-CNN models are limited,and they cannot extract discriminative features,resulting in a large number of pixels being incorrectly divided.However,the CNNSPGRU model requires a small number of samples,so it still maintains high classification accuracy under small sample conditions.