Abstract:
Abstract: Rice is one of the most important food crops in China and even the world. Effective monitoring of rice is essential to ensure national food security, effective water resources management, and greenhouse gas emissions. In recent years, remote sensing technology has been increasingly applied to rice information extraction. To find remote sensing data sources suitable for rice extraction in the mountain area and classification features suitable for each growth stage of the rice, and to compare the classification accuracy differences of rice at different growth stages. The study was based on Sentinel-2 images to perform the extraction. The data included Sentinel-2 image data, ASTER GDEMV2 elevation data, study area boundary vector data, and field survey data. Based on the characteristics of the growth stages of rice, we selected the seeding period from May to June, the growth period from July to August, and the maturity period from the end of August to September. Land-use type was divided into 5 categories: rice, other vegetation (dryland crops, grassland, weed wasteland), forest land (various woody plant land), water bodies, construction land (including buildings, roads, bare Ground). After pre-processing such as atmospheric correction and re-sampling, 58 classification features such as band features, vegetation index, red-edge index, water body index, topographical features, and texture features were selected. Classified samples were selected through outdoor field surveys and indoor visual interpretation. 682 rice, 557 other vegetation, 335 water bodies, 458 forest lands, and 514 construction land were collected. The mean and standard deviation were extracted and the classification features were screened using the SEaTH algorithm. After screening, the number of classification features was 16 during seeding, 13 during growth, and 12 during maturity. The data of each period were classified by the random forest classification method, and an error matrix was constructed to compare the classification results. The user accuracy of each period was 0.93, 0.88, and 0.85; the mapping accuracy was 0.93, 0.96, and 0.93; the overall classification accuracy was 0.92, 0.92, and 0.91; the Kappa coefficient was 0.90, 0.90, and 0.88, respectively. The study drew the following conclusions that rice was easily confused with different land types at different growth stages. The separation between various ground objects was better during the seeding period. It was easy to be confused with forest land and other vegetation during the growth period. It was easily confused with construction land and other vegetation during the maturity period. The classification features could be screened by SEaTH algorithm to avoid data redundancy. Band characteristics and vegetation index were the main classification features in each period. The accuracy of the user during the rice seeding period was the highest, and it was the best period for rice extraction. The method provided a feasible mapping scheme for remote sensing extraction of rice cultivation areas in the mountain area. This study was only applicable to small-scale rice extraction. It was bound to increase work costs and reduce efficiency for large-scale rice extraction. GEE (Google Earth Engine) as a remote sensing big data cloud platform, relying on its strong cloud computing capabilities to expand the scope of the research area in the future; the user accuracy obtained during the rice growth period was 0.88, which also achieved high classification accuracy. Could the image development method be used to combine the data from the seeding period with the growth period to obtain higher classification accuracy? The accuracy of classification needed further study in the future. The classification method used in this research was the pixel-based image classification. It was difficult to avoided false pixels. In the future, images could be segmented and the object-oriented classification method was used to eliminate the influence of false pixels.