Monitoring in-season biomass nitrogen in diversified cropping systems: a machine learning approach

Monitoring nitrogen (N) status in spatially diversified cropping systems is essential, but challenged by soil heterogeneity and crop-specific growth. This study assesses the feasibility of utilizing machine-learning models to estimate above-ground biomass and nitrogen (N) uptake across six co-occurring crops (barley, rye, wheat, rapeseed, maize and sunflower) grown in a patch-cropping arrangement. Random Forest and Extreme Gradient Boosting models were trained on multisource data collected from 2021 to 2024, integrating spectral indices, phenological information, soil and weather variables. We aimed to test whether shared spectral-phenological patterns enable cross-sectional estimation across different crops. Random Forest without soil variables achieved the highest performance (R2 = 0.75 for biomass and 0.68 for N uptake). BBCH code, near-infrared (NIR) reflectance, photosynthetically active radiation (PAR) and precipitation emerged as the most relevant predictors. The effect of soil inputs was algorithm-dependent: Random Forest achieved highest performance without soil variables, whereas XGBoost benefited from their inclusion. Crop-specific accuracy varied, with wheat performing best and maize and sunflower showing a tendency towards regression to the mean. These results demonstrate the potential application of unified N-monitoring frameworks in diversified cropping systems, while underscoring the need for further validation across regions and varying growing conditions for a diverse crop portfolio.