Maize yield forecasting in Lombardy region in Italy using a machine learning model driven by remote sensing data

Regional crop production is increasingly affected by climate variability, creating a need for operational monitoring and early-warning systems based on Earth observation (EO). In this study, we present an end-to-end EO-driven framework for regional maize monitoring in Lombardy (Northern Italy), combining annual maize mapping (2017–2025) with early-season yield forecasting over the same period.Maize distribution is mapped annually at 10 m resolution using Sentinel-2 imagery. A LightGBM classifier is trained on phenology-based NDVI features derived from seasonal composites. Training data are obtained from official crop-type raster maps of the Lombardy Regional Agricultural Information System and supplemented with provincial parcel data for 2024. To reduce commission errors, classification is restricted to cropland using the DUSAF “arable land” mask provided by Regione Lombardia.Maize yield forecasting relies exclusively on early-season information defined in thermal time (GDD < 1200). Field-level features are extracted by GDD stages from multiple EO and meteorological sources, including Sentinel-2 L2A spectral indices, Sentinel-1 GRD VV/VH backscatter, MODIS land surface temperature, evapotranspiration (ET/PET), and LAI/FPAR, ERA5-Land daily temperature, precipitation, radiation and soil moisture (with vapor pressure deficit derived), SMAP surface and root-zone soil moisture, and static terrain and soil properties from NASADEM and SoilGrids.A stacking ensemble model (Random Forest, Gradient Boosted Decision Trees, and XGBoost with a ridge regression meta-learner) is trained on an independent field-level maize yield dataset from Spain, linearly calibrated, and transferred to Lombardy. Regional and provincial yield estimates are further bias-corrected using standardized early-season anomaly features and an independent drought indicator (PDSI). When evaluated against official Lombardy maize yield statistics (7-province average), the anomaly- and PDSI-based correction substantially improves interannual performance, reducing RMSE from 1.20 to 0.53 t ha⁻¹ and increasing explained variance to R² ≈ 0.73.Overall, the proposed framework shows how phenology-based crop mapping and early-season, multi-source EO information can be integrated into a practical regional system for maize monitoring and yield forecasting, supporting climate risk assessment and adaptation planning.