Machine learning-assisted model for predicting biochar efficiency in colloidal phosphorus immobilisation in agricultural soils

The loss of colloidal phosphorus (Pcoll) from agricultural lands significantly contributes to nonpoint source nutrient pollution of receiving waters. This study aimed to develop an advanced machine learning (ML) model to predict the immobilisation efficiency of Pcoll (IE-Pcoll) by biochar in agricultural soils. Six ML algorithms were evaluated using a dataset containing 18 biochar- and soil-related variables. The random forest (RF) algorithm outperformed the others (R2 = 0.936-0.964, RMSE = 2.536-3.367), achieving superior test performance (R2 = 0.971, RMSE = 2.276). Key biochar-related parameters, such as oxygen content, total phosphorus content, and application rate were found to be stronger drivers of IE-Pcoll than most soil parameters. Soil Olsen-P was found to be a more reliable predictor of IE-Pcoll than the other soil-related parameters. Feature selection techniques narrowed down the original 18 features to the most critical ones, enhancing the performance of the model. A graphical user interface based on the optimised model was developed to provide practical field-based predictions of IE-Pcoll under varying conditions. This study highlights the strong potential of using biochar as a sustainable soil amendment to enhance Pcoll immobilisation, thereby reducing non-point source nutrient pollution from agricultural soils.