Study On Land Cover Classification Method Of Temporal Sentinel-2 Remote Sensing Data

Land cover changes naturally over time and is the result of human activity.Remote sensing technology has the ability of macroscopic,rapid and large-scale monitoring.It can obtain objective and accurate land cover information in a short time,and land cover products that meet a series of scientific and policy information needs are regularly produced on different space and time scales.Remote sensing technology has the advantages of wide range of data acquisition and rich information.Land cover classification based on remote sensing data has become one of the main methods of land cover classification.However,such methods have the problems that the image is disturbed by external factors such as clouds and fog,the data is "missing" in the time series,the fragmentation of the plot is large,and the planting structure is complicated,which leads to low classification accuracy.In view of the advantages of Sentinel-2 remote sensing data with high temporal and spatial resolution and easy access,this paper establishes a Sentinel-2 time series data set,uses a random forest classification method to construct a fine-grained land cover classification model based on multiple features,and uses cases The study discusses method efficiency,accuracy,and scalability.Experiments verify the feasibility of this method and the space-time expansion ability of the classification model under typical category system.The main work and results of the paper are as follows:(1)Realization of fine land classification in Xiongxian research area based on time series Sentinel-2 dataBased on the spectral band information(RGBN)and vegetation index(NDVI)classification features of the time-series Sentinel-2 image data,for the level＿2 in the classification system,a random forest classification algorithm is used to extract 12 types of land cover information in Xiongxian and surrounding areas for classification and extraction,and obtain the land cover distribution map of the study area.(2)Analysis of application of land cover classification model based on time-series Sentinel-2 image in time dimensionFor the three features of R-I,R-P,and R-I-P for image data in 2018 and 2019,it analyzes the extension and application of typical types of farmland,woodland,sparse forest,buildings,water bodies,and barren grassland in the time dimension.Among them,the classification accuracy of the three classification features in 2018 is higher than 79.09%.The Kappa coefficient is at least 0.6985,and the 2019 Kappa coefficient is higher than 83.87%.The results show that,based on the random forest algorithm and current survey sample data,time series data can be used to effectively extract land cover classification information for adjacent historical years.(3)Analysis of Extended Application of Classification Model Based on Time Series Features in Space SceneBased on the land cover classification model of the Xiongxian study area of the time series data,the land cover classification extension application and effectiveness evaluation of the T50 SMJ and T50 TMK images were carried out at the Beijing-Tianjin-Hebei junction.The scalability of the classification model in T50 SMJ is generally better than T50 TMK.Among them,the scalability based on R-I classification features is the best.The classification accuracy of R-I-P and R-P features is not much different.The overall classification accuracy is 88.90% and 86.33%,respectively.At the same time,the classification accuracy shows that the space expansion of farmland is the best,the expansion of forest land,buildings,and water bodies is the second,and the expansion of sparse forests and barren grasslands is weak.