Nitrous Oxide Emission Prediction by Combining Process-based Models and Neural Networks

Nitrous oxide (N2O) is a potent greenhouse gas (GHG) that has about 273 times more global warming potential over a $\mathbf{1 0 0}$-year time horizon than carbon dioxide. Legislations put in place by nation-states are pressuring farmers to reduce their GHG emissions, but verifying whether GHG reductions strategies are working requires affordable $\mathbf{N}_{2} \mathbf{O}$ emission sensing/estimation. Process-based models and data-driven models can be combined to offer an affordable $\mathbf{N}_{2} \mathbf{O}$ sensing solution for making reasonably accurate estimates. In the present work we gathered soil, weather, and N2O emissions data using infield sensors from a test plot at a Canadian smart farm during the 2021, 2022, and 2024 growing seasons. We performed N2O prediction experiments that involved predicting 2022 emissions by using models trained using 2024 and/or 2021 data. We compared the hybrid long short-term memory (LSTM) - Dynamic Land Ecosystem Model (DLEM) model (LSTM-DLEM) to other datadriven models. When using only 2021 training data, LSTMDLEM was the best performing model, achieving $0.39 \mathbf{R}^{2}$. When using different features and including 2024 training data, LSTM was the best performing model, achieving $0.40 \mathrm{R}^{2}$, and LSTMDLEM did poorly, although LSTM-DLEM was the only model that best captured the large scale of the emissions at the very start of the season. Our results show that the idea of combining data-driven models with process-based models has the potential to improve $\mathbf{N}_{2} \mathbf{O}$ emission prediction model performance.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.