Research on machine learning prediction of biochar yield based on “geographic three-dimensional information”

Biochar has high application value and broad prospects. Using machine learning to explore the yield of biochar has become a hot topic. This study innovatively introduces geographical three-dimensional information (longitude, latitude, and altitude), and establishes machine learning models based on Light Gradient Boosting Machine (LightGBM), Deep Neural Network (DNN), and Categorical Boosting (CatBoost) to predict biochar yield, while comparing the prediction accuracy across eight different input combinations. Combined with Shapley additive interpretation (SHAP), partial dependence graph (PDP) and individual conditional expectation graph (ICE), the influence of geographical three-dimensional characteristics on the yield of biochar and the sample specificity were analyzed. The results show that the introduction of geographic three-dimensional information significantly improves the accuracy of the model. The R2 of LightGBM increased from 0.932 to 0.954, MAE decreased from 4.171 to 3.299, and RMSE decreased from 5.986 to 4.817. The importance of characteristics reveals that biochar yield exhibits obvious geographical dependence. Longitude has the most significant effect on the yield of biochar, and the yield increases significantly with the increase of longitude, while latitude and altitude have relatively weak effects. This study provides a new framework and reference for the prediction of biochar yield and the utilization of biomass resources.