Mapping of residual soil nitrate in a diversified cropping system using multi-source sensing data and stacked ensemble learning

Abstract:

Spatial mapping of residual soil nitrate (NO3-) after harvest is essential for targeted nitrogen (N) management and reducing leaching risk. Specifically, predicting NO3- in spatially diversified cropping systems is challenging because soil properties, crop types, and fertilization rates vary between and adjacent crops. This study tested a stacked ensemble machine learning model to predict crop-scale post-harvest NO3- of the topsoil (0-30 cm) in a diversified patch cropping system in Brandenburg, Germany.

Five years of data (2020–2024) from 30 patches and 9 different crops were used. The framework integrated data on volumetric soil moisture, precipitation, Sentinel-2 vegetation indices, and N management variables aggregated over temporal windows from 1 to 90 days before harvest. The stacked ensemble of Random Forest and XGBoost achieved the highest performance for the 90-day window (RMSE = 11.5 NO3-N kg ha-1, MAE = 9.0 NO3-N kg ha-1, R2 = 0.61), outperforming both base learners individually. The feature importance analysis (SHAP) identified cumulative precipitation and NDVI as the dominant predictors. Predicted responses showed increased residual NO3- under low precipitation (<150 mm) and high N fertilizer rates (>150 kg N ha-1), whereas NO3- declined with high NDVI averages above 0.55. The five-year dataset captured both a persistent texture-driven spatial gradient in residual NO3- and a drought-related reversal in 2022, when sandy low-yielding patches exhibited unusually high post-harvest NO3- under exceptionally dry pre-harvest conditions. These results demonstrate that a multi-source machine learning approach can identify relative post-harvest NO3- hotspots in spatially heterogeneous diversified cropping systems with contrasting crop phenologies and management conditions. The framework provides a locally calibratable template for targeted N monitoring in comparable climates, spatially heterogeneous diversified cropping systems with the availability of soil, local weather, crop, and management data.

Read the full article here: https://www.sciencedirect.com/science/article/pii/S2949911926000675?via%3Dihub