Implementazione di modelli GIS per valutare l’idoneità colturale delle varietà di grano duro a diverse scale spaziali.

This study aims to develop yield prediction models for durum wheat using datasets that are easily accessible to agro-industrial companies. A curated database of 243 field records, collected between 2017 and 2023, integrated pedo-climatic, satellite, and varietal information. Meteorological and remote sensing variables were aggregated by phenological stages estimated through a Growing Degree Day model, while the MSAVI-2 index was derived from Sentinel-2 imagery. Given the limited sample size and high multicollinearity among predictors, both regularization methods (Ridge, Lasso, Random Forest) and a stepwise regression with pre-selection of non-collinear variables were applied. Stepwise regression yielded the most parsimonious and interpretable models (R² ≈ 0.66; RMSE ≈ 0.8 t ha⁻¹), consistently highlighting wind and rainfall during tillering, thermal differences in later stages, and soil silt content as key predictors. The MSAVI-2 index at booting stage emerged as a strong early predictor, enabling forecasts about one month before harvest. However, satellite data alone proved insufficient without cultivar or pedo-climatic information. Overall, the research demonstrates the potential of combining remote sensing and pedo-climatic data for reliable yield prediction, while emphasizing the need for standardized and expanded data collection to enhance model robustness.