A General Model for Large-Scale Paddy Rice Mapping by Combining Biological Characteristics, Deep Learning, and Multisource Remote Sensing Data

Due to the influences combined with global climate change and human activity, paddy rice area and distribution have undergone dramatic changes. Currently, many approaches for paddy rice mapping rely on the prior knowledge of paddy rice phenology or require widely distributed ground samples of paddy rice, which are limited for large-scale applications. In this work, we propose a general paddy rice mapping (GPRM) model by combining biological characteristics, deep learning, and multisource remote sensing data. The proposed GPRM first utilizes the normalized difference vegetation index and land surface water index to acquire large-scale remote sensing dataset in key phenology periods of paddy rice, such as the transplanting period and peak vegetative growth period. Then, a general model using object-based deep neural networks is developed and trained by the remote sensing dataset and the ground reference data collected in one region (e.g., Guangdong Province), which can be directly applied for generating 10-m paddy rice maps in other regions with different climate conditions and complex cropping systems (e.g., Jiangxi Province and Heilongjiang Province). The results demonstrate that the GPRM can realize remarkable performance of paddy rice mapping in China. The overall accuracies are over 99%, and the user accuracy, producer accuracy, and Kappa coefficient vary from 0.77 to 0.93, 0.94 to 0.97, 0.9 to 0.95, respectively. Overall, the GPRM is has significant promise for large-scale paddy rice mapping with complex cropping systems, thus supporting global agricultural development strategies and food security.