Knowledge-guided machine learning captures key mechanistic pathways for better predicting spatio-temporal patterns of growing season N2O emissions in the U.S. Midwest

Abstract

Accurately predicting agricultural N₂O emission hot moments has long been a key focus of N₂O models, which is challenging given the complex mechanisms involved and the high spatio-temporal heterogeneity of controlling factors. In this study, we improve a knowledge-guided machine learning model for agricultural N₂O flux prediction (KGML-ag-N₂O) by (i) incorporating a physical module to explicitly represent fertilization, (ii) pre-training KGML-ag-N₂O with synthetic data from a process-based (PB) model ecosys under diverse fertilization strategies and environmental conditions, and (iii) fine-tuning KGML-ag-N₂O with field observations of daily N₂O flux and key controlling factors from 29 sites across the U.S. Midwest. We then assess whether it can outperform the PB model and the pure machine learning model over different spatio-temporal scales. Through integrating knowledge from both the PB model and field observations, KGML-ag-N₂O shows the best performance in capturing site-level daily N₂O emission hot moments (r = 0.63, RMSE = 3.765 mgN m^-2 d^-1), mainly due to the strengthened triggering effect of soil water content increase on N₂O emissions in KGML-ag-N₂O. The improved causality representations between key controlling factors and N₂O emissions further lead to KGML-ag-N₂O outperformance in capturing N₂O emission patterns over larger spatial (e.g., regional) and temporal (e.g., inter-monthly) scales than both conventional approaches. By validating and interpreting the improvements in KGML-ag-N₂O performance, our study illustrates its potential to quantify agricultural N₂O emissions in the U.S. Midwest, and provides important insights into identifying and reducing PB model structure uncertainties using KGML techniques.