Research On Remote Sensing Supervision Classification Method Based On Sub-Sample Analysis

With the in-depth study of high-resolution remote sensing technology in China,the remote sensing related industries are booming,and domestic high-scoring satellite products are widely used in geology,environment,transportation and many other fields.How to use remote sensing to obtain typical elements and their change information quickly,with high quality and automatically becomes an urgent problem to be solved.The basis of most remote sensing applications is the classification and information extraction of remote sensing images.The accuracy of image extraction depends on the selection of reference samples.The existing sample selection and image interpretation methods are still artificial,and the automatic extraction has been in the research stage.Technically,the automatic selection of samples and the implementation of classifications need to be further strengthened;From the results,the classification results are affected by the problems of “homologous” and “homogeneous foreign matter”,which makes the automatic extraction of remote sensing fall into the bottleneck.For the difference in image characteristics of the same species,this paper proposes a remote sensing supervision classification method based on subsample analysis.The concept of "subsample" is proposed,that is,the sample is established by subclass as the standard category,and "subclass" refers to different categories with different image features within the same feature type.This paper selects the remote sensing image of GF-2 from the north slope of the Tianshan Mountains in Xinjiang as experimental data,and realizes the automatic extraction based on subsamples for the saline-alkali land information in the study area.The research uses dynamic clustering algorithm to automatically obtain subclasses,and then complete the establishment of subsample libraries.In order to test the applicability of the subsamples,three sets of contrast experiments were designed by using the control variable method.The Kappa coefficient and the overall classification accuracy were used to measure the extraction effect between the large sample and the subclass samples and between the three supervised classifiers.In order to realize automatic extraction,the classification method of convolutional neural network was used to complete the extraction experiment efficiently and accurately,and the convolutional neural network method was compared with the traditional extraction method.The main results are:(1)It realizes the automatic division of subclasses and the automatic acquisition and construction of subsamples,which has certain guiding significance for the research of supervised classification based on pixels;(2)Through contrast experiments,it is proved that the remote sensing classification method based on sub-sample analysis effectively improves the accuracy of remote sensing information extraction and weakens the influence of isomerism to a certain extent;(3)Using the convolutional neural network to realize the automatic extraction of remote sensing information,saving time and manpower,and improving the efficiency and accuracy of classification.