Land Cover Classification Of Typical Yellow River Source Regions Based On Deep Learning

Land use/cover change(LUCC)is closely related to climate change,human activities,surface hydrological and ecological processes,etc.Land cover information is the basic data for the study of regional climate change response,hydrological and ecological evolution and land use change.Research on land cover information extraction methods based on high-precision satellite images and deep learning can improve the efficiency,accuracy and intelligence of land cover classification,which is of great significance for rapid and quantitative acquisition of land cover types and characteristics.Taking Three River Source National Park-Huangheyuan Park as the object,this paper adopts the deep learning method to carry out the research on LUCC,so as to explore the classification and recognition effect of various deep learning semantic segmentation methods in the Huangheyuan Park,which is dominated by alpine meadow.The main work of this paper includes:selecting appropriate data sources,developing land cover classification system,processing remote sensing image data,selecting spectral features that can reflect the difference of land features,and constructing sample data set;Based on the FCN8S model,the batch standardization process was added to construct the FCN8S?BN model.Based on FCN8S?BN,SEGNET and PSPNET,different sample cutting schemes were tested to determine the optimal sample size and image overlap rate.The classification performance of FCN8S and FCN8S?BN models was compared,and the applicability of each semantic segmentation method to the alpine ecosystem in the Huangheyuan Park was explored by comparing the classification results of SEGNET,PSPNET,HRNET and UNET.The FCN8S?BN model was used to extract multi-year water body and analyze the change of water body area.The main results and conclusions are as follows:(1)The sample size and the overlap rate between adjacent samples have a certain influence on the classification effect.Landsat 8 OLI images were used as data sources,and near-infrared,red and short-wave infrared 1 band images were selected as RGB composite images for the experiment.The results show that increasing the cut size of samples can improve the classification accuracy of ground features,especially the ground features with large size.When the overlap rate increases from 0 to 50%,the classification accuracy is improved obviously.When increasing from 50%to 75%,the accuracy of forestland and grassland with high coverage increased obviously.For the Huangheyuan Park,the cut size of 128×128 and the overlap rate of adjacent image blocks is 75%are selected to build a good sample data set.(2)The classification experiments based on different semantic segmentation models show that each precision index of FCN8S?BN model is higher than that of FCN8S,and the pixel accuracy is 81%,16.6%higher than that of FCN8S.Compared with the five semantic segmentation models,FCN8S?BN is slightly better than the other models,among which UNET model has the worst effect,unable to identify several categories with less sample data.(3)Huangheyan Park of Three River Source National Park has a water area of about 1,486?1,590 km~2 in recent 30 years.Since 2000,the water area has been expanding,with an average annual increase of 10.12 km~2.Precipitation is positively correlated with water area,and its influence on water area is greater than evaporation.