Nitrous oxide (N2O) emission from soils results from microbial nitrification and denitrification processes and is regulated by environmental conditions. N2O emission is enhanced when there is a surplus of inorganic nitrogen in the soil, such as after applying nitrogen fertilizer. Various temporal and spatial conditions can lead to these microbial processes in soils and result in emissions that have been characterized as occurring in “hot spots” and at “hot moments”.
Due to these characteristics, it is challenging to predict emissions using existing approaches and, therefore, to track progress in mitigation efforts through improved nitrogen management practices. The current approach to predict N2O emissions is based on emission factors or on process-based (PB) soil biogeochemical models. PB models are more suitable to predict “hot spots” and “hot moments” than emission factors but have limitations due to incomplete knowledge of processes leading to N2O emissions. In contrast, machine learning (ML) models learn patterns and relationships from data but require large amounts of data for training and are not easily interpreted.
To address these challenges, Liu et al. (2022) developed a first-of-its-kind knowledge-guided machine learning model for agroecosystems (KGML-ag) by incorporating knowledge from an advanced PB model, ecosys. The authors trained the KGML-ag model on “synthetic data” generated by ecosys; about 4.3 million data points (122 days by 18 years by 99 mid-western US counties by 20 nitrogen fertilizer rates) were partitioned into training, validation, and testing. The KGML-ag model was tested with observed daily N2O fluxes from controlled-environment experiments and always outperformed the PB and ML models in predicting N2O flux, capturing its temporal dynamics and emission peaks.
Compared to ML models, KGML models require significantly less data since they use data generated by the PB model for pre-training. However, further improvements will require high temporal resolution N2O observation data from field measurements. As global efforts aim to better monitor N2O fluxes, the approach by Liu et al. (2022) shows promise in improving future prediction of N2O fluxes.
