Abstract
Ammonia (NH(3)) is a major pollutant in poultry farms, negatively impacting bird health and welfare. High NH(3) levels can cause poor weight gain, inefficient feed conversion, reduced viability, and financial losses in the poultry industry. Therefore, accurate estimation of NH(3) concentration is crucial for environmental protection and human and animal health. Three widely used machine learning (ML) algorithms-extreme learning machine (ELM), k-nearest neighbor (KNN), and random forest (RF)-were initially used as base algorithms. The wavelet transform (WT) with ten levels of decomposition was then applied as a preprocessing method. Three statistical metrics, including the mean absolute error (MAE) and the correlation coefficient (R), were used to evaluate the predictive accuracies of algorithms. The results indicate that the RF algorithms perform robustly individually and in combination with the WT. The RF-WT algorithm performed best using the air temperature, relative humidity, and air velocity inputs with a MAE of 0.548 ppm and an R of 0.976 for the testing dataset. In summary, applying WT to the inputs significantly improved the predictive power of the ML algorithms, especially for inputs that initially had a low correlation with the NH(3) values.