Hyperspectral Remote Sensing Image Classification Based On CNN And MRF

Hyperspectral remote sensing images contain abundant information of ground objects,and the methods for interpretation and classification of hyperspectral remote sensing data have always been the focus of research.In this paper,a hyperspectral remote sensing image classification method based on convolutional neural networks and Markov random field is studied for the characteristics of high latitudes and training data limitation of hyperspectral remote sensing images.An improved convolutional neural network based hyperspectral remote sensing image classification method with spatial-spectral feature combination is studied.This method redesigned the structure of traditional convolutional neural network,uses some special convolution kernels which have the same dimensions with spectral bands to extract the deep combined feature of training data,this method adds Dropout control vector between all fully connected layers and use the ReLU functions as its activation functions,proposed a convolutional neural network with its deep feature extraction ability and strong generalization ability.Compared with other methods in the international remote sensing experimental data set,the results show that SSCF-CNN has good accuracy and can classify different types of data sets with high precision.In view of the "salt and pepper" phenomenon in the HSI classification results,a hyperspectral remote sensing image classification algorithm based on convolutional neural network and Markov random field is constructed by combining this SSCF-CNN model with Markov random field.The algorithm constructs the labelled field based on the classification results of SSCF-CNN,builds the SSCF-CNN-MRF framework based on the labelled field and the real data sets,and completes the optimization of the classification results of the SSCF-CNN.In view of the "isotropy" problem in the HSI classification,an ensemble learning hyperspectral remote sensing image classification algorithm based on Bagging and multi wheel voting was constructed.This method amplifies the training samples and constructs multiple training subsets and randomly assigned them to base classifiers.Finally,the classification results are fused with the previous classification results,and the classification accuracy is improved and the edge optimization problems were solved.