Pre-harvest forecasting of rainfed wheat yield in Iran using multi-source remote sensing and machine learning

Accurate prediction of rainfed wheat yield in arid and semi-arid regions such as Iran is essential for food security under increasing climate variability. This study developed and compared machine learning (ML) models integrating multi-source remote sensing and meteorological data to forecast rainfed wheat yield across 22 Iranian provinces (2001–2021). We utilized MODIS-derived vegetation and climate indices (NDVI, LAI, LST, FPAR, ET), TerraClimate’s PDSI drought index, ERA5 precipitation and temperature, and FLDAS soil moisture at two depths. Three ML algorithms: Random Forest, XGBoost, and Support Vector Regression were trained using varied training-test splits (70–90 %) and forecast windows (2–4 months pre-harvest). Hyperparameters were optimized through grid and randomized search. XGBoost with 80 % training data (n = 352) achieved superior accuracy (R 2 =0.64, mean-normalized NRMSE=42 %, MAE=0.25 t/ha), outperforming RF (R 2 =0.60) and SVR (R 2 =0.55). April (2 months pre-harvest) yielded optimal predictions, with key predictors identified as NDVI (April, November, December), precipitation (February, March), LST (April, December), and LAI (April). These results enable supplemental irrigation scheduling and precision agriculture in water-limited environments. They demonstrate the operational value of integrated remote sensing and ML for agricultural planning in arid and semi-arid regions. • XGBoost outperformed RF and SVR in rainfed wheat yield prediction. • Model accuracy peaked with 80 % training data and two months before harvest. • NDVI, precipitation, LST and LAI were key predictors of wheat yield. • Rainfall and soil moisture during early growth stages strongly influenced yield. • Integrated approach enables early warning for wheat production in different climates.