Machine learning-based prediction of N2O emissions in coffee soils under different nitrogen fertilization regimes

Jibaja J.A.F.; Troyes J.F.O.; Saavedra W.L.D.; Ojeda J.M.P.; Julcapoma M.R.; Vasquez K.L.D.; Castillo A.E.H.; Vázquez G.G.; Pino M.E.M. · Clean Energy Science and Technology (2026) · DOI: 10.18686/cest595

Resumen

Nitrous oxide (N2O) emissions from nitrogen-fertilized agricultural soils represent a critical yet poorly quantified component of the greenhouse gas balance in Peruvian coffee systems. This study developed and compared machine learning (ML) and multiple linear regression models for predicting N2O fluxes from coffee soils under five contrasting nitrogen fertilization treatments in Chirinos, San Ignacio, Cajamarca, Peru. A dataset of 135 observations (n = 135), integrating edaphic and meteorological variables, was assembled from field measurements using static closed chambers and gas chromatography. Models were trained on 80% of the data and validated on the remaining 20% using holdout cross-validation. Random Tree and Random Forest consistently outperformed traditional regression approaches, both achieving R2 = 0.94, whereas REPTree reached R2 = 0.44, and multiple linear regression showed limited predictive capacity (R2 ≤ 0.50). The best-performing model, Random Tree Model IV, achieved a mean absolute error (MAE) of 14.79 mg N2O–N m−2 d−1 . Variable importance analysis identified nitrogen source, soil temperature, relative air humidity, porosity, and water-filled pore space as the primary drivers of emission variability, consistent with established microbial mechanisms of nitrification and denitrification. These findings demonstrate that ML-based approaches substantially outperform conventional linear methods for modeling the nonlinear dynamics of N2O in coffee soils and provide actionable evidence for nitrogen management strategies that prioritize organic sources as mitigation options in high-rainfall environments. Limitations include the one-month evaluation period and the absence of microbiological variables, which future studies should address to improve model transferability.

Instituciones: Universidad Nacional de Ingeniería; Universidad Nacional de Jaén