Monthly 1-km yield mapping without census statistics: An approach combining satellite, crop model, field data and machine learning tested in three countries with diverse rice cropping patterns

抄録

　Gridded yield products are essential for understanding the historical trajectories of food production, assessing climate risks in cropping systems, validating tools for generating yield projections, and contributing to developing a knowledgebase for making agri-food systems sustainable. This study aims to develop a method for deriving a monthly 1-km gridded yield product that is theoretically applicable worldwide and without national or subnational yield statistics. The method combines objective sources of information, i.e., satellite remote sensing, process model simulation and field measurements, via a machine learning residual model approach and does not rely on yield statistics. The method was tested for feasibility in Madagascar, Sri Lanka and the Philippines, where the number of annual rice harvests are different. The gridded products derived using the proposed method produce yield estimates that are consistent with the independent data. However, the consistency varies by season, satellite, and data type (crop cuts, household surveys, yield statistics and other gridded yield products). The 1-km yield estimates agreed better with the field observations than with the yield statistics and surveys. The relative root-mean-squared errors (rRMSE) appeared smaller for the wet-season than for the dry-season. The yield estimates based on MODIS and VIIRS are relatively the most reliable, followed distantly by those derived using GCOM-C. The yield products developed here are unique in that they provide information on season-specific water management conditions and yield responses to changes in nitrogen input and planting dates, providing guidance for adaptation to climate change through agronomic adjustments.