Modeling ammonia concentration in swine building using biophysical data and machine learning algorithms.

• Factors affecting NH 3 concentration in swine building are not fully understood. • Monitoring and quantifying NH 3 are crucial for maintaining barn environment. • Three machine learning algorithms were utilized to predict NH 3 concentration. • SVR had a better performance accuracy compared to other tested models. • Pigs' body mass and feed intake mostly influenced the NH 3 concentration. Ammonia (NH 3) concentration in livestock barns is a crucial environmental parameter affecting the well-being of animals and workers' health. Effective management and prediction of NH 3 levels are essential for maintaining an efficient and enduring swine production system. This study investigated the application of machine learning algorithms, namely support vector regression (SVR), random forest regression (RFR), and multiple linear regression (MLR), to predict NH 3 concentrations in pig barns and examined the impact of individual input variables on prediction accuracy. In this study, three datasets, each comprising five key biophysical variables, i.e., feed intake (FI), mass of pig (MP), carbon dioxide (CO 2) levels, temperature (T), and relative humidity (RH), were utilized for training and testing these algorithms. The data were collected from three barns during the growing-finishing stage of pigs in 2022 and 2023. The study results revealed a strong positive relationship between FI and MP with NH 3 concentrations. Among the three machine learning models, the SVR outperformed the MLR and RFR in predicting NH 3 concentration. The result exhibited that the SVR obtained the maximum performance in both training (R2 >0.95) and testing (R2 >0.85), with R2 improvements of up to 5.43 % and 14.02 % and RMSE decreases of up to 15.97 % and 28.98 %, in comparison to the RFR and MLR across the three input datasets in NH 3 prediction. The study also emphasized the importance of dataset size, with the large dataset containing all five input indicators achieving the highest accuracy compared to smaller datasets. In addition, the MLR demonstrated maximum stability, followed by the SVR, whereas the RFR exhibited minimum stability. Sensitivity analysis revealed that FI was the most influential input variable for NH 3 concentration prediction. The study ranked the impact of individual input variables as FI > MP > CO 2 > T > RH. The combination of FI, MP, CO 2 , and T as input indicators achieved the highest model performance, accounting for a substantial portion of the variance between observed and predicted data. This study demonstrated the potential of machine learning models, particularly SVR, for predicting NH 3 concentrations using relevant input variables in pig barns. These findings significantly enhance understanding of NH 3 concentration dynamics in pig barns, providing crucial insights into swine production and environmental management with data-driven modeling. [ABSTRACT FROM AUTHOR]