High-resolution wall-to-wall time series of seasonal maize area and yield for Rwanda over 2019–2023

Abstract Agricultural monitoring is least developed for smallholders in low- and middle-income countries—communities most likely to be impacted by hunger, poverty, and climate change. Recent efforts to monitor smallholder productivity are limited in spatial and temporal scope, but here, we provide an end-to-end machine learning pipeline built on Google Earth Engine for high-resolution, wall-to-wall time series mapping of crop area and yield, demonstrated for maize at every 10 m pixel in Rwanda over 2019–2023. Gradient boosted tree models were built from more than 60 000 field-level labels, 9000 yield measurements, and satellite-derived inputs. Maize was classified with 83% accuracy, precision of 0.70, and recall of 0.44 and total maize cover was predicted within 4% of national statistics. Yields aggregated to districts had an RMSE of 370 kg ha −1 (nRMSE: 27%). Our data compare favorably to other smallholder maize classification and yield estimation products for sub-Saharan Africa while being accessible, low-cost, standardized, and observed over time; thus, being more likely to enable technology transfer and downstream analyses.