Objectives and Methods With the development of high-precision and intelligent seismic exploration techniques, it is necessary to accurately quantify and interpret coal seam thickness to provide reliable data for constructing high-precision three-dimensional geological models. This will help ensure safe and efficient coal mining in coal mines. In coal seam thickness prediction, the comprehensive multi-attribute analysis method faces challenges such as the cumbersome selection of optimal seismic attributes and the low predictive accuracy of regression models. To address these challenges, this study proposed a support vector regression (SVR) prediction model based on kernel principal component analysis (KPCA) and particle swarm optimization (PSO) (also referred to as the KPCA-PSO-SVR model). This model employed KPCA for dimensionality reduction to reduce redundant information in seismic attributes. By combining the PSO algorithm, this model automatically optimized the hyperparameters (penalty coefficient c and kernel function parameter γ) of the SVR model, thereby overcoming the limitations of empirical parameter adjustments and effectively improving the prediction accuracy and efficiency. Furthermore, the introduction of an incremental learning mechanism, integrated with the dynamic incorporation of actual data revealed by mining faces, enabled the model to achieve dynamic quantitative interpretation of the coal seam thickness.

Results and Conclusions  Numerical analysis of the wedge-shaped forward model demonstrated that the dimensionality reduction via KPCA optimized multi-attribute combinations, providing high-quality data for the regression analysis model. The quantitative comparison of the prediction results of different models indicates that the KPCA-PSO-SVR model yielded a coefficient of determination (R²) of 0.9667, outperforming traditional SVR, PSO-SVR, grey wolf optimizer (GWO)-SVR, and back propagation neural network (BP-Net) models. Both the KPCA-PSO-SVR and PSO-SVR models exhibited high prediction accuracy and computational efficiency under varying proportions of training datasets. Field application to a survey area demonstrates that the KPCA-PSO-SVR model displayed excellent robustness and generalization capability under a small sample size in the coal mining area. This model yielded an R² value of 0.9858, higher than the values of its counterparts. After incremental training was conducted by incorporating actual data revealed by roadways of the mining face into the model, the maximum absolute error of the predicted thickness decreased from 0.622 m to 0.178 m, further enhancing the model’s adaptability and predictive performance. The KPCA-PSO-SVR model provides an effective technical solution for coal seam thickness interpretation, holding great significance for engineering applications.