Heavy metals bioavailability prediction in swine manure composting: Development of ensemble machine learning models

Assessing the bioavailability of heavy metals (HMs) in compost products is crucial for evaluating the related environmental problems. For this, machine learning (ML) models have shown an excellent accuracy in HM fraction prediction and solve the limitations of experimental methods. Thus, this study aimed to predict the bioavailability of different HMs (Cu, Zn, Cd, As, and Cr) in swine manure compost using AdaBoost, random forest (RF), CatBoost, K-nearest neighbors (KNN), and extreme gradient boost (XGB) algorithms. Afterward, three multi-model ensemble techniques, namely AdaBoost ensemble (ABE), Neuro-Fuzzy ensemble (NFE), and weighted average ensemble (WAE), were developed using the outputs of single ML models as input to boost the overall prediction accuracy. The prediction accuracy of the proposed models was evaluated using statistical indices, namely determination coefficient (R²), mean absolute error (MAE), Nash Sutcliffe Efficiency (NSE), and root mean square error (RMSE), and graphical illustrations. The prediction results demonstrated that the best single models for Cd, Cu, Zn, As, and Cr were AdaBoost (NSE = 0.941), XGB (NSE = 0.918), AdaBoost (NSE = 0.959), AdaBoost (NSE = 0.973), and KNN (NSE = 0.859), respectively. The results of multi-model ensemble techniques indicated that the nonlinear NFE gave the highest accuracy, improving the accuracy of individual models for Cu, Cd, As, and Cr by 4.139 %–10.265 %, 5.314 %–26.564 %, 2.78 %–7.197 %, and 4.79 %–24.238 %, respectively based on testing set NSE values.