Modeling nitrogen optimization and yield prediction in durum wheat using multi-source sensor, climate, and plant trait data through advanced machine learning algorithms

This study provides a comprehensive evaluation of sensor-based measurements and artificial intelligence-supported modeling approaches for yield and grain quality prediction in durum wheat (Triticum durum) grown under semi-arid conditions. Optical sensor data (SPAD, NSPAD, NDVI, INSEY) and plant height measurements were collected across four phenological stages (ZD24, ZD30, ZD31, ZD32), along with RGB and multispectral UAV imagery, climatic variables, and laboratory-based spectral reflectance data covering the 350–2500 nm range. The effects of different nitrogen application rates on physiological, morphological, yield, and quality parameters were examined using analysis of variance (ANOVA), which revealed significant increases in SPAD, NDVI, and INSEY values and notable differences in yield and quality components with increasing nitrogen supply. Five machine learning algorithms (Random Forest, Gradient Boosting, AdaBoost, LightGBM, XGBoost) were comparatively evaluated for yield prediction. Model performance varied across growth stages, with the highest prediction accuracy observed at ZD31, corresponding to the critical phase of active biomass accumulation and nitrogen assimilation. LightGBM and XGBoost exhibited superior performance with lower error metrics and higher generalization capacity. SHAP (SHapley Additive exPlanations) analysis identified INSEY and NSPAD as the most influential predictors in yield estimation. Additionally, a CNN-based deep learning model integrating time-series UAV imagery, ground-based sensor measurements, and climatic parameters demonstrated high predictive accuracy (R² ≈ 0.98) and improved performance particularly during heading and grain-filling stages. For protein prediction, Random Forest, Deep Learning, hybrid ANN–SVR, and PLSR models were tested across four spectral regions (VIS, NIR, SWIR, Full Spectrum). The hybrid ANN–SVR model using full-spectrum and SWIR data outperformed others, achieving the highest accuracy (R² > 0.92) and the lowest error values. Overall, the findings highlight the multidimensional impact of nitrogen management and emphasize that concurrent evaluation of physiological, morphological, yield, and quality traits is essential for precision fertilization strategies. The study demonstrates that sensor-based data fusion combined with machine and deep learning approaches offers a powerful, explainable, and scalable decision-support framework for sustainable wheat production and precision agriculture applications.