Research On Key Techniques Of Remote Sensing Image Classification Based On Deep Learning

Remote sensing image classification is the key technology of geographic information system,which plays a very important role in the field of urban planning,environmental monitoring,military reconnaissance and so on.In the field of machine learning,deep learning is springing up in recent years.By mimicking the hierarchical structure of hu-man brain,deep learning can extract features from lower level to higher level gradually,and distill the spatio-temporal regularizes of input data,thus improve the classification performance.However,there is little work in the present work to apply the deep learning methods to the remote sensing image classification task.The main work and innovation of this paper are as follows:(1)A classification method of SAR remote sensing images based on the deep belief network(DBN)model is proposed.DBN combines the advantages of unsupervised learn-ing and supervised learning,and it can automatically extract the spatio-temporal features of remote sensing data,and then improve the classification accuracy.The experimen-t results show that the proposed method can achieve better classification performance than the support vector machine(SVM)and the traditional neural network(NN)by using multi-date RADARSAT-2 satellite polarimetric SAR image.(2)A new method of hyperspectral remote sensing image classification based on hi-erarchical local receptive field extreme learning machine(ELM)is proposed.The LRF concept originates from research in neuroscience.Considering the local correlations of spectral features,it is promising to improve the performance of HSI classification by in-troducing the LRFs;Besides,the hierarchical structure can further improve the classifica-tion accuracy of hyperspectral image.Experiments on Indian Pines and Pavia University hyperspectral data sets show that our proposed method can achieve higher classification accuracy compared with other methods such as SVM,and has faster training efficiency.(3)A new method of hyperspectral remote sensing image classification based on second order moment sparse coding is proposed.Sparse coding(sparse representation)is applied to hyperspectral remote sensing image processing,and it is a hot research direc-tion in the field of hyperspectral information processing in recent years.However,these methods based on sparse coding do not make full use of the spatial information of hyper-spectral remote sensing image.By mining the spatial correlation of the information,it is promising to further enhance the classification accuracy of hyperspectral remote sensing image.Experiments on Tiangong-1 and KSC data sets were conducted and the resutls show the proposed method can greatly improve the classification accuracy of hyperspec-tral remote sensing image.(4)An optical remote sensing image scene classification method based on hierarchi-cal sparse coding is proposed.This method is essentially a kind of multi-layer,multi-scale and multi-path sparse coding.It can extract features of optical remote sensing images more effectively,so that the features of the remote sensing images can be represented more sufficiently.The obtained codes are further used for Spatial Pyramid pool(SPP)operation,and the corresponding SPP representation is obtained.SPP representations in different paths are combined and outputted to the SVM classifier,and the final classifi-cation results are obtained.Experiments on 3 data sets(UCM,WHU-RS and RSSCN7)show that the method proposed in this paper can get a higher scene classification accuracy of the optical remote sensing images.