The stacking method enhances machine learning models for monitoring and understanding regional soil available nitrogen variations in croplands

Abstract

Soil available nitrogen (AN) is crucial for crop growth, grain yield, and sustainable agricultural management. The Guanzhong Plain (GP) is an important grain production area in the Yellow River basin of China with intensive agricultural activities for over 2000 years and excess nitrate loading. To predict the spatial distribution of AN in the root zone (0–100 cm), 124 soil samples were collected via borehole drilling, followed by lab analysis and AN prediction model development (machine learning models, MLMs and ensemble models, EMs). The results indicated that nitrate (NO₃^--N) and ammonia (NH₄⁺-N) contents declined with increasing depth, with significantly higher values in the upper 40 cm. NH₄⁺-N contents were lower and relatively stable across soil layers. EMs outperformed MLMs, with the stacking method performing better and improving averaged R², RMSE, and MAE by 10.48 %, 4.93 %, and 6.99 % for NO₃^--N prediction and 6.75 %, 9.41 %, and 8.94 % for NH₄⁺-N prediction. Soil variables were most critical for NO₃⁻-N prediction, contributing 46 % of the relative importance, followed by topography (22 %) and climate (17 %). NH₄⁺-N predictors were dominated by topographic variables, accounting for 51 %. These findings highlight the distinct roles of soil and topography in regulating nitrogen dynamics, with soil properties controlling nitrification and leaching processes for NO₃⁻-N and topography influencing water redistribution and retention for NH₄⁺-N. This study provides references to precise fertilizer management and non-point source pollution control in GP. It also underscores the potential of ensemble models, particularly stacking, in improving AN prediction accuracy across agroecosystems.