Research On Classification Of Hyperspectral Image Based On Deep Learning

The classification of hyperspectral remote sensing images is a hot research issue in re-cent years.The characteristics of quantified continuous spectral curves and high spectral resolution of hyperspectral images provide conditions for their classification.However,the hyperspectral data has a large number of bands,high correlation between the bands,large spectral dimensions,and easy to cause excessive information,which makes the research of hyperspectral image classification face some challenges.Therefore,solving the problem of information redundancy of hyperspectral data and effectively extracting the deep features of hyperspectral data are the keys to solving hyperspectral image classification.In this pa-per,two types of deep learning models-convolutional neural network model with supervised learning and autoencoder model with unsupervised learning are used to discuss the classifi-cation of hyperspectral images.Secondly,the depth of the deep learning model determines the effect of training the model.If the model layer is too shallow,it may cause underfitting,and if the model is too deep,it may cause overfitting.In order to effectively extract the deep features of hyper-spectral data,the classic 2D deep convolutional neural network is reconstructed to obtain a new deep learning network model that is based on convolutional neural network,and the model was trained on the commonly used Indiana Pine Forest dataset and the University of Pavia dataset to extract the deep features of the image.The comparative analysis of the experimental results verified the superiority of the model.Secondly,the depth of the deep learning model determines the effect of the training model.If the model layer is too shallow,it may lead to under-fitting.If the model is too deep,it may lead to over-fitting.For this problem,through reconstructing the classical 2D depth convolutional neural network,a hyperspectral image classification model based on convolutional neural network was proposed.A new deep learning network model suitable for hyperspectral image classification was obtained.The contrast experiments were carried out on two commonly used data sets,Indian and PaviaU for the superiority of this verification model.Finally,manual labeling of sample points in hyperspectral images takes time and effort,so it is necessary to study the classification of hyperspectral images with a small number of labeled samples.Aiming at this problem,a semi-supervised learning-based hyperspectral image classification model based on autoencoder is proposed,which uses a combination of unsupervised learning and supervised learning for network training on hyperspectral data.Model training was performed on the Indiana Pine Forest dataset and the University of Pavia dataset to verify the validity of the model.