Machine learning for ammonia volatilization prediction and slurry application management

Abstract

Anthropogenic ammonia emissions primarily originate from agriculture, especially field fertilization. These emissions represent nitrogen loss for farmers and contribute to air pollution, posing risks to human health and the environment. Estimating ammonia emissions is crucial for national inventories and policy-making. Various models exist for predicting emissions, including mechanistic, empirical, and semi-empirical approaches. While machine learning (ML) is widely used in environmental science, its application to ammonia emissions remains limited. In this study, we used 5939 ammonia emission data from 538 trials, extracted from the ALFAM2 database, to train three machine learning methods - random forest, gradient boosting, and lasso - for predicting cumulative ammonia emissions 72 h after manure application. These methods were compared to the semi-empirical ALFAM2 model using an independent test dataset. Random forest (RMSE = 4.51, r = 0.94, MAE = 3.28, Bias = 0.92) and gradient boosting (RMSE = 6.19, r = 0.89, MAE = 4.10, Bias = 0.51) showed the best performance, while the lasso log-linear model (RMSE = 7.30, r = 0.84, MAE = 5.57, Bias = -1.38) performed worst. Both random forest and gradient boosting outperformed the semi-empirical ALFAM2 model, which showed performance comparable to the lasso model. We then used these models and the ALFAM2 model to compare five slurry management techniques, varying in application method (trailing hoses, trailing shoes, and open slot) and post-application incorporation, across 128 scenarios with different manure types and weather conditions. Compared to broadcast application, alternative techniques reduced emissions by a median of -13.6 % to -61.7 %. This study highlights the promise of ML models in assessing ammonia emission reduction methods, while emphasizing the importance of evaluating model sensitivity to algorithm choice.