Application of Machine Learning Algorithms in Nitrous Oxide (N2O) Emission Estimation in Data-Sparse Agricultural Landscapes

To understand if machine learning algorithms could be employed in agricultural landscapes to estimate N2O emissions, multiple linear regression (MLR), random forest regression (RFR), support vector regression (SVR) and artificial neural network (ANN) algorithms are tested on an agricultural site in Ontario, Canada. Two scenarios, High Input (HI) and Low Input (LI), were used to check the performance of these algorithms’ using R2, RMSE, VE, p-factor, r-factor and visual inspection indicators. The HI consisted of discrete measurements of N2O, rainfall, temperature, fertilizer application dates, soil nitrate, ammonium content and pH values, whereas the LI scenario did not use the latter three. The results indicated that MLR was inapplicable as the data did not satisfy its fundamental assumptions. RFR, SVR and ANN under HI were able to capture 64% (66%), 59% (63%) and 94% (43%) of the variability of emissions within the training (testing) datasets. Subsequently, these models were able to capture 92%, 29% and 75% of high emissions (>10 gm/ha/day) within their predictive intervals of 95% confidence. RFR, SVR and ANN under the LI scenario captured 72% (68%), 61% (66%) and 81% (68%) of the variability in N2O emissions within the training (testing) datasets. While these models were found to have comparable performance in both HI and LI scenarios, HI was found to be better at capturing high emissions. Based on the computational cost, ease in finetuning, capture of peak emissions and stable model performance, RFR and ANN are recommended to estimate N2O emissions in the study area and similar agricultural landscapes in future studies.