Mapping paddy rice with the random forest algorithm using MODIS and SMAP time series

Rice production is very important for national food security in China. Time series vegetation indices and phenology-based algorithms have been utilized to map paddy rice fields by identifying the flooding and seedling transplanting phases from multitemporal moderate-resolution (500 m to 1 km) images. Satellite imagery in other electromagnetic spectrum such as microwave region provides supplementary information about land surface characteristics such as soil moisture, which may improve the accuracy of rice mapping. In this study, we developed a method for identifying paddy rice at the regional scale based on the potential relationship between soil moisture and crop growth. Random forest classification optimized by the coordinate descent algorithm was used to generate a 500 m resolution paddy rice map with time series downscaled soil moisture based on Soil Moisture Active Passive (SMAP) images and Moderate Resolution Imaging Spectroradiometer (MODIS) images of the northern Songnen Plain in Northeast China-one of the major paddy rice cultivation regions in China. A combination of the Enhanced Vegetation Index (EVI), Land Surface Water Index (LSWI), Phenological Parameters (PH), and Soil Moisture (SM) was used to identify the rice fields during the flooding/transplanting and ripening phases. Two scenarios (EVI + L SWI + PH and EVI + L SWI + PH + SM as the input parameters) were developed to extract the paddy rice from the images. A comparison of the two identification scenarios indicated that the addition of the soil moisture to the combination achieved higher identification accuracy, especially at the junction of the different crop phases. The proposed method could be beneficial to researchers attempting to improve the accuracy of paddy rice identification at regional scale. (c) 2020 Published by Elsevier Ltd.