Résumé
This study addresses the critical need for accurate reference evapotranspiration (ET0) estimation in data-scarce environments such as West Africa, where rapid population growth and climate change intensify water resource challenges. Traditionally, computing ET0 with the FAO-56 Penman-Monteith method requires multiple meteorological inputs—such as solar radiation, humidity, and wind speed—collected from costly, fully instrumented weather stations. However, the availability and maintenance of such equipment can be prohibitive in remote regions. To overcome these constraints, we explore the potential of machine learning (ML), specifically XGBoost (XGB) models, trained on NASA Power climate reanalysis datasets. Our approach relies on a limited subset of easily measured in situ variables—daily minimum and maximum temperatures and rainfall—to estimate ET0. Departing from standard ML practices that depend on short-term, site-specific data, we leverage the extensive historical depth and broad spatial coverage of reanalysis products. We trained and validated twenty locally adapted XGB models using measurements from twenty diverse West African weather stations. Our results show that certain XGB model configurations, notably those incorporating temperature and rainfall data, can approximate ET0 estimates from the FAO-56 Penman-Monteith equation with median RMSE values frequently below 1 mm/day—levels comparable to commonly employed empirical formulas. This finding demonstrates that minimal on-site instrumentation, combined with ML and reanalysis data, can effectively support irrigation scheduling and enhance water-use efficiency under varying agro-ecological conditions. To foster broader implementation, we have released our XGB-ET0 code under the GPLv3 licence (https://github.com/SARRA-cropmodels/RF-ET0), enabling researchers and practitioners to locally train and deploy these models, thereby improving ET0 estimation and sustainable agricultural water management in resource-limited settings.