Mapping irrigation regimes in Chinese paddy lands through multi-source data assimilation

Water-saving irrigation (WI) is a crucial agricultural management with the benefits to save irrigation water, reduce energy consumption, and suppress methane emissions from paddy lands. Classifying WI practices from traditional flooding irrigation (FI) is a key component in detecting the rice irrigation status, which is significant to estimate the total agriculture-associated greenhouse gas emissions. In this study, we developed an automatic method to map irrigation regimes across Chinese paddy lands. First, we used seven variables related with irrigation facility or vegetation cover as proxy to generate representative WI and FI samples. Besides, we composited 123 features of optical bands and synthetic aperture radar from MODIS and Sentinel-1 data. Then, we trained a random forest model for each province with these samples. Finally, we applied the trained model to generate maps of WI/FI practices at 500 m resolution. Comparisons of the resultant maps with census data indicated highly accurate estimations of the WI area at a city- or province-level, with a R2 higher than 0.92. The overall accuracy of the classification was approximately 0.73, as validated through ground truth samples. Additionally, we also conducted a data quality analysis and confirmed the classification results were reliable in main rice production area of China. With the push towards carbon neutrality goals and the increasing demand for clean management practices, we developed and demonstrated an advanced method to produce near real-time maps of irrigation regimes and provide crucial data support for agricultural emissions reduction and irrigation management decisions.