Research On Hyperspectral Image Classification Algorithm Based On Neural Network

Since the hyperspectral image itself contains a large amount of spectral information and spatial information,it can use each pixel of the hyperspectral image to identify and classify the feature.Therefore,the hyperspectral remote sensing image technology has received widespread attention from researchers in recent years..In order to obtain more abstract and deep-level feature information,deep learning methods can be used to construct a reasonable and effective neural network for feature extraction.The thesis mainly uses 3D convolutional joint attention network and multi-feature learning method to construct a network to classify hyperspectral images,and then uses the adversarial neural network to classify hyperspectral images.The main contents are as follows:First,the 3D convolutional joint attention mechanism network is used to classify hyperspectral images.The network can select features that are more critical to the current task target from a large number of features in the feature learning process to make the classification effect better.Through comparative experiments,it is found that this method is effective in classification of hyperspectral images.Second,using different attribute filters to extract different features,and using multi-feature learning to extract more representative deep features,can effectively classify hyperspectral images.This method can help to extract a better feature representation through the learning of each convolutional layer,let various features show their own characteristics,and use the heterogeneity between different features to enhance feature extraction.Finally predict the final label of each hyperspectral image pixel.The method is analyzed and compared through comparative experiments,which shows that the relevant classification model can improve the accuracy of hyperspectral image classification.Third,a classification model based on generative adversarial neural network is designed.The model can use the extracted spatial features as input data of discriminator D together with the original hyperspectral image data,and use a 3D convolutional neural network as the discriminator D's.The network can capture key features that are more important for the classification of hyperspectral images during the training process,which is helpful for thegenerator G to generate "fake data" with excellent performance,and make the "fake data" as close to the real training samples as possible.And it can learn the spatial and spectral features in the hyperspectral image at the same time,in order to reduce the impact of spectral variability on classification.Find out the advantages and disadvantages of this method through comparative experiments.