Observational modeling study demonstrates accurate biomass and nitrogen prediction across diversified patch crops, suggesting viable unified agricultural monitoring frameworks.
Key Points
To assess whether machine learning models can accurately estimate above-ground biomass and nitrogen uptake across six co-occurring crop species using shared spectral and phenological indicators.
Trained Random Forest and Extreme Gradient Boosting models using multisource data gathered from 2021 to 2024 across six patch-cropped species (barley, rye, wheat, rapeseed, maize, and sunflower).
Integrated vegetation spectral indices, phenological metrics (BBCH scale), weather records, and soil variables to evaluate model transferability across diverse crops.
Random Forest models excluding soil inputs yielded the highest overall predictive performance, achieving an R² of 0.75 for above-ground biomass and 0.68 for nitrogen uptake.
BBCH phenological stage, near-infrared reflectance, photosynthetically active radiation, and precipitation were identified as the strongest predictors across models.
Algorithm sensitivity to soil features differed, with XGBoost improving when soil inputs were added, while crop-specific accuracy peaked in wheat and showed regression toward the mean in maize and sunflower.