Abstract: Geometry-based simultaneous localization and mapping (SLAM) methods struggle to meet the accuracy requirements of roadway cross-section extraction, thereby limiting the reliability of deformation monitoring and safety assessment. Underground roadways are typical geometrically degenerate environments, where the scarcity of distinguishable geometric features poses significant challenges to accurate mapping and cross-sectional analysis. To address these issues, this study proposes a LiDAR SLAM method integrating geometric features with adaptive intensity features, together with a centerline-based cross-section extraction method. Adaptive intensity information is incorporated into geometric feature extraction to enhance edge feature discrimination, while an adaptive weighting strategy is developed to dynamically adjust the contributions of geometric and intensity features according to underground environmental conditions. A high-quality global point cloud map is generated through robust state estimation and subsequently projected onto the XOY plane. Centerline points are extracted using a neighborhood-shift iterative strategy, and normal slices are generated along the centerline based on a prescribed thickness threshold, enabling refined geometric characterization of roadway cross-sections. The results demonstrate that the proposed SLAM method achieves stable loop closure across different testing environments. Even in challenging mining roadways, it maintains high localization accuracy and robustness, with a maximum translational error, root mean square error, and mean error of 0.312, 0.172, and 0.163 m, respectively, in terms of absolute trajectory error. By projecting the digital model onto a two-dimensional plane, the proposed method directly extracts discrete centerline points, with errors ranging from 0.002 to 0.027 m, demonstrating high accuracy and stability. Furthermore, considering the irregular contours of mining roadways, the reconstructed cross-sections effectively capture local geometric details, achieving a maximum fitting overlap ratio of 99.07%. These findings demonstrate the strong geometric characterization capability of the proposed method and provide reliable data support for intelligent coal mining.