Sourface Features' Information Extraction Of High-resolution Remote Sensing Image And Application

With the launch and application of China's Gao Fen 2(GF-2)remote sensing satellite,the high-resolution remote sensing image(HRRSI)has increased sharply.The extraction of sourface features' information from a large number of HRRSIs without annotation information is a key issue for HRRSI interpretation.However,a HRRSI not only covers a range of thousands of square kilometers,but also presents a completely different set structure and spatial pattern in different regions,and using traditional computer vision methods to solve the problem of HRRSI sourface features' information extraction can not meet the accuracy of engineering applications.Therefore,this paper proposes a method for extracting HRRSI sourface features' information based on scene classification.The basic idea is to divide the HRRSI into sub-regions of different scene classes by scene classification method,then extract the sourface features' information based on the HRRSI sub-region of the specific scene,and an efficient and robust sourface features' information extraction prototype system is developed.The main work of this paper is as follows:(1)To divide HRRSI into sub-areas of different scene classes,a remote sensing image scene classification(RSISC)method which is hybrid deep convolutional neural network(CNN)tranfer learning based on NASNet and Restricted Boltzmann machine(RBM)is proposed.The method uses the RBM network instead of the fully connected layer of the NASNet network when transferring the pre-trained NASNet network to the target dataset,and fine-tunes to retrain the RBM network and the softmax classifier layer on the target dataset.Experiments on the datasets of AID,NWPU-RESISC45,UC-Merced,WHU-RS19 and RSSCN7,the hybrid deep CNN obtains the highest classification accuracy of 95.98%,91.30%,98.56%,97.44% and 94.57%,respectively.After that,it is applied to the HRSSI scene classification of the study area,and accurate scene division results can be obtained.(2)To realize the sourface features' segmentation of different scene sub-regions of HRRSI,an unsupervised object segmentation method(DPMM-OMRF model)based on Dir ichlet process mixture model(DPMM)and Markov random field(MRF)is proposed.Firstly,the super-pixel was divided into basic objects by mesh.Secondly,DPMM prior was constructed by multidimens ional Gaussian distribution,and MRF prior was constructed by similar ity measure;prior distribution of DPMM-OMRF model was thus calculated by adjusting the two priors via adaptive weight.Subsequently,the DPMM-OMRF model was built by the likelihood distribution and joint prior distribution under Bayes ian framework,and the conditional distribution of class labels was deduced.Finally,to update the label f ield and parameters of the DPMM-OMRF model,a Gibbs sampling method was designed by deriving and calculating the class label posterior probability.The experimental results show that the overall accuracy of the DPMM-OMRF model reaches up to about 90%,and the Kappa coefficient is close to 0.8,and the total number of sourface features' classes can be identified,and sourface features can be segmented more accurately and completely.After that,it is applied to the sourface features' segmentation of HRRSI sub-regions in different scenes of the study area,and all show excellent performance.(3)Combining with the hybrid deep CNN is used to RSISC and DPMM-OMRF model is used to realize the unsupervised object segmentation,and HRRSI sourface features' information extraction prototype system is designed and implemented.The scene classification module and the unsupervised segmentation module are respectively implemented by using the ENVI secondary development tool IDL language;and the human-computer interaction module is developed to realize the interactive extraction of various types of sourface features' information.The HRRSI of the study area is input into the system,and various types of sourface features' information in the HRRSI can be accurately extracted.