Research On Classification Technology Of Hyperspectral Remote Sensing Image Based On Convolutional Neural Network

Hyperspectral remote sensing images have been widely used in the fields of land resources and military security due to the characteristics of unification of atlas and massive information.Using hyperspectral remote sensing images to classify ground objects is the key task of its application.In the classification of hyperspectral remote sensing images,it is difficult to obtain a large number of labeled samples because data labeling not only needs to be performed pixel by pixel,but also requires field investigation to determine the category of each pixel.Therefore,under the condition of limited labeled samples,high-precision classification is one of the main research contents in the field of hyperspectral remote sensing research.At present,deep learning methods have also been widely used in hyperspectral image classification.Among them,the classification method based on convolutional neural network can achieve better classification results.However,due to the information redundancy in hyperspectral images and the limited number of labeled samples,the performance of the algorithm suffers in practical use.This research has carried out the following research on the above problems:(1)This research proposes an end-to-end attentional two-channel residual network based hyperspectral remote sensing image classification method.When the training labeled samples are limited,the classification accuracy becomes low due to insufficient redundant information and spectral-spatial feature extraction in hyperspectral images.This method takes the original data cube as input data,and firstly uses the spectral-band attention module to The influence of spectral band redundancy and spatial interference pixels on classification is reduced;then spectral and spatial bi-branch networks are used to fully extract spectral and spatial features in hyperspectral images,in which Mish activation function is used.A series of experiments on three hyperspectral datasets,Indian pines,University of Pavia and Salinas Valley,show that the proposed method achieves good classification results under the condition of limited training samples.(2)This research proposes a semi-supervised classification method of hyperspectral remote sensing images based on improved ladder network is proposed.Due to the problem of low classification accuracy due to limited labeled samples,this method uses both labeled samples and unlabeled samples to train the network model.To solve the problem of "dimension disaster" caused by direct use of raw data,principal component analysis algorithm is used to reduce data dimensionality;At the same time,in order to make full use of spectral and spatial information in HSI,a hybrid spectral convolutional network is proposed for feature extraction.A series of experiments on two hyperspectral datasets,the University of Pavia and Salinas Valley,show that the proposed algorithm achieves better classification performance and shows better practical value under the condition of using a small number of labeled samples and some unlabeled samples.(3)This research designs a hyperspectral remote sensing image classification system.Based on the above classification algorithm,a hyperspectral remote sensing image classification system is developed.The system includes an image loading module,an image preprocessing module,an image classification module and a performance evaluation module.It not only realizes the functions of image visualization and data dimensionality reduction,but also realizes the functions of hyperspectral image classification method selection,related parameter setting and classification result evaluation.