Downscaling dekadal WaPOR actual evapotranspiration from 300 m to 20 m for irrigation monitoring

Abstract Irrigated agriculture accounts for nearly 70% of global freshwater withdrawals, yet monitoring consumptive water use at field scale remains challenging in fragmented smallholder systems. FAO’s WaPOR database provides global dekadal actual evapotranspiration (ETa) at approximately 300 m resolution, but this scale cannot resolve sub-field heterogeneity typical of southern Africa. This study presents WaPOR-AETI-DownScaler v1.0, an operational framework for downscaling WaPOR ETa from 300 m to 20 m while assessing whether large-scale magnitude consistency is maintained. The approach applies CatBoost regression using Sentinel-2 reflectance and vegetation indices (NDVI, NDMI, FVC), harmonic seasonal encoding, CHIRPS rainfall, terrain variables, and coarse-resolution WaPOR ETa as a structural predictor. Model development over two WaPOR Level-3 pilot areas in Mozambique used approximately 500,000 dekadal pixel samples and 150 feature-subset optimization trials. Internal validation, used for model selection, achieved R² = 0.883, RMSE = 3.95 mm dekad⁻¹, and RRMSE = 19.5%. Independent temporal testing on hold-out dekads, used for generalization assessment, maintained stable pixel-level performance (mean pvn_r² ≈ 0.62; relative RMSE ≈ 23–24%). No statistically significant difference was detected between aggregated 20 m predictions and WaPOR Level-1 ETa at 300 m (p > 0.05), indicating no detectable degradation of aggregated 300 m ETa magnitude after downscaling. The framework enhances fine-scale ETa heterogeneity without detectable degradation of basin-scale magnitude, providing a scalable tool for irrigation monitoring and water accounting in data-scarce regions.