Machine learning reveals dynamic controls of soil nitrous oxide emissions from diverse long-term cropping systems.

Soil nitrous oxide (NO) emissions exhibit high variability in intensively managed cropping systems, which challenges our ability to understand their complex interactions with controlling factors. We leveraged 17 years (2003-2019) of measurements at the Kellogg Biological Station Long-Term Ecological Research (LTER)/Long-Term Agroecosystem Research (LTAR) site to better understand the controls of NO emissions in four corn-soybean-winter wheat rotations employing conventional, no-till, reduced input, and biologically based/organic inputs. We used a random forest machine learning model to predict daily NO fluxes, trained separately for each system with 70% of observations, using variables such as crop species, daily air temperature, cumulative 2-day precipitation, water-filled pore space, and soil nitrate and ammonium concentrations. The model explained 29%-42% of daily NO flux variability in the test data, with greater predictability for the corn phase in each system. The long-term rotations showed different controlling factors and threshold conditions influencing NO emissions. In the conventional system, the model identified ammonium (>15 kg N ha) and daily air temperature (>23°C) as the most influential variables; in the no-till system, climate variables such as precipitation and air temperature were important variables. In low-input and organic systems, where red clover (Trifolium repens L.; before corn) and cereal rye (Secale cereale L.; before soybean) cover crops were integrated, nitrate was the predominant predictor of NO emissions, followed by precipitation and air temperature. In low-input and biologically based systems, red clover residues increased soil nitrogen availability to influence NO emissions. Long-term data facilitated machine learning for predicting NO emissions in response to differential controls and threshold responses to management, environmental, and biogeochemical drivers.