| Since the 21 st century,with the continuous development of science and technology,hyperspectral remote sensing image classification technology,as the basis of remote sensing image technology,has become an increasingly popular research topic.Hyperspectral remote sensing classification technology has been successfully applied in many fields such as natural disaster monitoring,environmental protection,military and national defense,and has shown a very important practical role.Many researchers apply cutting-edge technologies to the field of hyperspectral image classification,trying to find breakthroughs.Compared with traditional multispectral image data,hyperspectral image data has more bands,stronger correlation between bands,and stronger ability to distinguish spatial and spectral features.However,when extracting feature information from remote sensing images,traditional classification methods often face problems such as poor generalization ability,large amount of calculation,and long time,and the classification accuracy is also difficult to meet the standard.In addition,conventional remote sensing datasets have small sample features,and it is difficult,costly and time-consuming to collect the true labels of features.At the same time,classification using a small number of labeled training samples faces many challenges.In recent years,deep learning based on multi-dimensional convolutional neural networks has achieved great success in extracting deep features of data,but there are still problems of small samples and insufficient feature extraction.Therefore,in order to improve the above problems,this paper focuses on the classification of hyperspectral remote sensing images combined with multi-dimensional convolutional neural networks.The main tasks of the paper are as follows:1.This paper systematically studies the existing hyperspectral remote sensing image classification technologies,including typical support vector machines and their extended models based on this algorithm,and studies the popular multi-dimensional convolutional neural network methods,including GAN,1D-CNN,2D-3DCNN,3D-CNN and other algorithms,introduced the basic principles of the algorithms,and summarized the advantages and disadvantages of the algorithms in different scenarios,based on which the problems in the existing models were improved.2.Aiming at the problem that the samples generated by the existing multi-dimensional convolutional neural network-based GAN algorithm and the original samples have a large gap,the final classification accuracy of the classifier is low,and a GAN-based data enhancement model is proposed.This paper improves the existing GAN and adds a controller to solve the problem of low classification accuracy in the case of limited samples.At the same time,it can generate fake sample data that is closer to the original sample distribution,so that the GAN algorithm can learn the original sample data through the training process of the entire network.The characteristics of the samples enable the improved GAN classification model to generate more data samples that are very close to the original images.Comparative experiments show that the performance of the improved GAN network model is significantly better than other existing models,and the classification accuracy is effectively improved.3.Aiming at the problems of too many parameters and incomplete feature extraction in the existing multi-dimensional convolutional neural network 3D-CNN model,this paper proposes a hyperspectral remote sensing image classification algorithm based on 3D-CNN.A feature information enhancement model is added to the existing 3D-CNN network model,so that the network can better extract the spatial spectral features of the image,and can solve the problem of more irrelevant redundancy and more noise caused by information enhancement during the network training process.To solve the problem,the spatial filtering method is used to optimize the sample data after feature information enhancement,so that the enhanced sample data is smoother.Experiments show that the improved algorithm can more effectively and deeply extract the features of the original samples,and has higher classification accuracy. |
