Research On Hyperspectral Remote Sensed Image Classification Based On Convolutional Neural Network

Hyperspectral remote sensing images are rich in spatial information and spectral information,and they have been widely used in many practical scenarios.Hyperspectral remote sensing image classification is an important basic component of remote sensing image research and has received more and more attentions.The purpose of classifying every pixel in the image is to provide the ground-based information for follow-up remote sensing image research.This paper focuses on hyperspectral remote sensing image classification techniques based on convolutional neural networks.The work content of this paper is mainly divided into the following two parts:1)Classification model of feature fusion based on convolutional neural network:Taking into account the one-dimensional structure of the spectral dimension of hyperspectral remote sensing images and the two-dimensional structure of spatial geometry,two convolutional neural network models are designed to extract the pixels separately:Spectral information and spatial information.The spectral information extracted here is the original spectral dimension of the image obtained as the input of the first convolutional neural network model（spe-CNN）,and the extracted spatial information is obtained by reducing the dimension of the original spectrum of an image block of a certain size neighbourhood of each pixel according the second convolutional neural network model（spa-CNN）,and then the two extracted features are merged,this article uses concatation as the fusion methond,to make up for the deficiency of the two features in the spectral or spatial information.Finally,the support vector machine is used to classify the two features.This article refers to this classification model as CNN₂-SVM,and the classification accuracy after fusion can be higher than the single model spe-CNN or spa-CNN,but we found that the classification accuracy is very low based on the spe-CNN single model（we intuitively think the extracted features are not good enough）.Considering the noise in the spectral dimension of the image,this paper uses the denoising method to de-noise the original image first.Experiments show that the pre-processed method（PR^*-CNN₂-SVM）can Improve classification accuracy based on CNN₂-SVM.2)Classification model based on end-to-end convolutional neural network combined with post-processing:Since CNN₂-SVM divides feature extraction and classification into two steps,the classification process is more tedious.In order to further simplify the classification structure,this paper replaces SVM with softmax,so that the end-to-end training model is called CNNs in this paper,because PR^*proves the effectiveness of denoising in the PR^*-CNN₂-SVM classification method,so this method is also used in this model,but in the course of the experiment we found that the accuracy of training is higher than the accuracy of the test,and there is an over-fitting phenomenon.In response to this phenomenon,a post-processing method of weighted mean filter is proposed to deal with this phenomenon.Experiments have proved that PR^*-CNNs combined post-processing called（PR^*-CNNs-post）can obtain classification accuracy equivalent to PR^*-CNN₂-SVM under sufficient samples,and obtain higher classification accuracy than PR^*-CNN₂-SVM under few sample conditions.