Interpretable Machine Learning Identifies Nitrogen Turnover Signatures in Soil Ammonia Emissions

Soil ammonia (NH3) emissions are a major pathway for reactive nitrogen loss, but their spatial variability is difficult to explain using equilibrium-based approaches alone. This study evaluated whether machine learning can improve the interpretation of NH3 flux by identifying informative soil predictors. NH3 flux, soil temperature and volumetric water content (approximately 0–6 cm) were measured at 20 sampling points on bare agricultural soil, while NH4+, NO3−, total nitrogen, humus content and pH were determined at three soil depths. Five regression models were evaluated using leave-one-out cross-validation. Pairwise correlations showed that no single measured variable exhibited a dominant linear relationship with NH3 flux. Nevertheless, Random Forest achieved the best cross-validated performance (R2 = 0.418), suggesting that the observed spatial variability was better represented by multivariate and potentially nonlinear relationships among soil properties than by individual linear associations. Partial Least Squares Regression provided complementary information on the multivariate soil gradients. Within the Random Forest model, NO3− concentration at 10–20 cm and near-surface soil moisture had the highest permutation importance. These findings indicate that small-scale spatial variation in NH3 flux was associated with the combined statistical information contained in several soil properties rather than with a single dominant linear predictor. The importance of NO3− represents a dataset-specific statistical association and should not be interpreted as evidence of direct causal control or nitrification intensity.