Integrated Lasso-XGBoost-SVR Model for Subsurface Coal Seam Thickness Prediction

Abstract

Accurate prediction of coal seam thickness is a critical factor for enabling intelligent and efficient mining. To improve seismic-attribute-based coal seam thickness prediction under high-dimensional, redundant, and nonlinear conditions, this study develops a task-oriented Lasso-XGBoost-SVR hybrid framework. The proposed framework combines least absolute shrinkage and selection operator (Lasso) regression, extreme gradient boosting (XGBoost), and support vector regression (SVR). By integrating the complementary strengths of these three methods, the model identifies the key attributes controlling coal seam thickness. It also enables joint screening and coordinated fusion of heterogeneous data, reducing feature information loss introduced during sampling. This approach improves the accuracy of feature fusion and enhances predictive performance. On the independent test set, the proposed Lasso-XGBoost-SVR framework achieved a root mean square error (RMSE) of 0.0178, a mean absolute error (MAE) of 0.0105, and a coefficient of determination (R2) of 0.9732, demonstrating lower prediction errors and a better fit than the comparison models. Under the same data partition and evaluation protocol, the ablation study further showed that the integrated model consistently outperformed the Lasso-SVR and XGBoost-SVR variants across all three evaluation metrics. The model is also validated in a real mining area, where the predicted coal seam thickness closely matches geological observations. These results indicate that the Lasso-XGBoost hybrid approach is highly effective for optimizing seismic attributes in predicting coal seam thickness. This fusion strategy not only enhances predictive accuracy but also exhibits strong potential for practical implementation within mining panels, serving as a reliable tool for decision-making in operational mining scenarios.