Study Of Hyperspectral Image Classification Algorithm Based On Residual Network

With the vigorous development of hyperspectral remote sensing imaging technology,the massive acquisition of hyperspectral images have promoted its wider application.The hyperspectral image is a three-dimensional data cube that composed of one-dimensional spectral features and two-dimensional spatial features.Among them,the spectral features include rich features,while the spatial features include the detailed spatial distribution of features.Classification research is one of the main methods for mining rich information of images.Traditional classification methods based on hyperspectral images not only have low utilization of spatial and spectral information,but also often destroy the correlation between information in the process of feature extraction,and can't obtain ideal classification results.In recent years,with the continuous development of deep learning,more and more experts and scholars have applied it to the field of hyperspectral remote sensing,and based on this,a variety of hyperspectral image classification models have been proposed,such as convolutional neural network,deep belief network,etc.These classification models can automatically extract data features,reduce man-made operation processes,and also reduce the degree of damage to information in the feature extraction process,but with the continuous deepening of the network layer,the gradient disappears or disappears in the feature extraction process.The phenomenon of gradient explosion leads to the problem of overfitting,which directly affects the classification performance.Based on the above problems,this paper proposes three hyperspectral images classification models based on residual network,which are two-dimensional residual network composed of two-dimensional convolution kernel,three-dimensional residual network composed of three-dimensional convolution kernel and wide residual network perform classification research on hyperspectral images.Using two-dimensional residual network to classify the spectral feature information,spatial information and fusion spatial-spectral information of hyperspectral images respectively,the results show that two-dimensional residual network uses fusion spatial-spectral information to obtain the best classification results.Compared with two-dimensional residual network,three-dimensional residual network can use three-dimensional convolution kernels to directly extract the joint spatial-spectral information in hyperspectral images,which maximizes the correlation between the information and reduces the image data structure during feature extraction of destruction.This study combines three-dimensional residual network and virtual samples to construct a hyperspectral image classification model.After comparison with other classification models,it is found that this model not only alleviates the problem of insufficient training samples,but also obtains a good classification effect.However,because residual network pursues the network depth too much and ignores the problems of the model itself,the classification performance of the model has not been significantly improved with the increase of the number of residual units.Aiming at this problem,this paper proposes wide residual network based on hyperspectral images classification,and applies this model to carry out classification experiments on the spectral feature information,spatial information and fusion spatial-spectra information of hyperspectral images.The results show that the model has achieved excellent performance.Not only the training time is greatly reduced,but also the classification performance is excellent.It will have good development prospects in the field of hyperspectral image classification.