Application Of Visible Light And SAR Collaborative Ground Object Classification

Land resources are the material carrier on which human beings depend for survival and development,and are "the source of all production and all existence".To make rational use of land resources,it is necessary to identify and classify the ground objects first,and then further rationally plan the land resources on this basis.The exploration and research on the classification methods of ground objects at home and abroad have made some achievements in the classification of ground objects based on highresolution visible light.Because of the complex spectral information of ground objects,it is difficult to identify some ground objects.Polarimetric Synthetic Aperture SAR can acquire remote sensing data all day and all day,and it can record the microwave backscattering information of ground objects completely.Combining the complementarity of optical and SAR data,combining the band information of optical and polarimetric SAR remote sensing data,the land cover is classified and identified.In this paper,the highresolution visible sentinel No.2 data and Gaofen No.3 data are combined with multi-source data bands to classify the ground objects in the study area.The main work and research results of this paper are as follows:(1)the production of data sets,namely,Gaofen No.3 data with 3m resolution in Pengyang County and Sentinel No.2 data with open source in the same area provided by Ningxia Telemetry Institute.After the data is ready,under the SARscape plug-in of ENVI5.6,the remote sensing image data is preprocessed by data reading,focusing,multi-view processing,filtering,geocoding,calibration and final resampling.Then,the areas with rich ground features(including vegetation,buildings,water bodies,bare land,etc.)in Pengyang County are selected for clipping and data band information combination.In the preprocessing process,because the terrain of Pengyang County in Guyuan City is relatively high,DEM images of corresponding areas are needed in the pre-processing geocoding and calibration,and the corresponding DEM images are downloaded in the SARscape plug-in of ENVI5.6 in many inquiries and collections.Finally,the research area is cut and marked by deep learning method,which makes full preparations for subsequent experiments.(2)Research on the algorithm of full convolution neural network.The research content of this paper is the classification of ground objects based on pixel level.According to the characteristics of the model,FCN and U-net models are chosen for experiment and analysis.FCN model recovers the attribute categories of pixels from abstract features.Compared with CNN,CNN can extract more feature information.Based on this,this paper chooses FCN model for experiment.Because artificial data sets are limited,and U-net model can get good classification accuracy in the case of a small number of data sets,U-net model was chosen.Compared with FCN,the accuracy of U-net model in visible light data sources increased by 0.4%,in SAR data sources increased by 1.6%,and under the cooperation of two data sources,u-net improved by 1.1%.(3)Research on the improved full convolution neural network algorithm.In the experimental analysis of FCN and U-net methods,the classification accuracy is better,but the classification of details and edges is lacking.The processing of detail information is generally improved through channels,so SE module is proposed in this paper,which improves the detail information through attention mechanism to highlight the features of useful information,and then locates the useful information in the pool layer to improve the classification accuracy.Experiments show that U-net+SE model can improve the classification of ground objects.Compared with U-net and FCN model,the classification accuracy of visible light and SAR collaborative ground objects is improved by 5.1%and 6.2%,respectively.