Drivers of N2O Emissions: Implications for Model Development Accounting for the Spatial Variation

Cultivated soils contribute approximately 60% of global nitrous oxide (N₂O) production due to nitrogen inputs, which underscores the urgent need for comprehension of N₂O emissions at larger spatial and temporal scales. However, knowledge gaps persist due to the episodic nature of soil N₂O emissions, which are driven by non-linear interactions among biophysical, and environmental factors over spatial and temporal domains.

This study aims to identify significant predictors of N₂O emissions at the field scale using random forest algorithm. The soil N₂O flux and various predictors (CO₂, soil moisture content (SWC), temperature, mineral N, pH, bulk density, and air permeability, as well as digital elevation model (DEM), gamma ray count rate, and electrical conductivity data were measured between March and June 2024 in a 1.2 ha winter wheat field located in Foulumgård, Denmark. The N₂O flux was measured at 96 locations in weekly to biweekly time intervals using the LiCOR 7820 analyzer.

The N₂O flux spatially varied from 0.006 to 0.164 ug m^-2 s^-1, with the highest average fluxes of 0.148 ug m^-2 s^-1 approximately 7 to 10 days after fertilizer application. The CO₂ flux ranged from 0.11 to 0.54 µg m^-2 s^-1 with an average of 0.35 µg m^-2 s^-1, while SWC varied from 0.11 - 0.30 m³ m^-3  and soil temperature from 6.0 - 25.7 °C.

The preliminary random forest model identified key predictors for N₂O emissions as soil respiration (CO₂, 25%), temporal variability (Week, 13%), soil electrical conductivity, here a likely proxy for soil texture (EC, 11%), and SWC, 9%. Furthermore, the model was evaluated with a 90:10 data split, using 90% for training and 10% for validation. The absence of further predictors limited the model's performance, as reflected in the decline in R² from 89% in training to 60% in validation. The out-of-bag (OOB) error also showed the model explained only 29.4% of emission variability, emphasizing the need for additional variables to better capture N₂O predictors.

These findings are a first step towards comprehending the importance of recognizing the non-linear underlying forces of N₂O emissions and the intricate interplay between soil and environmental factors. Improving the model ability to predict N₂O emissions will require comprehensive datasets that capture key biogeochemical drivers and the development of robust, non-linear modeling frameworks. In a next step, additional parameters such as soil nitrate (NO₃^-), ammonium (NH₄⁺), and soil pH are included in the model to further improve model performance to understand spatial variation and temporal dynamic of N₂O.