3D point cloud lithology identification based on stratigraphically constrained continuous clustering

Three-dimensional laser scanning provides high-precision spatial data for automated lithology identification in geological outcrops. However, existing methods exhibit limited performance in transition zones with blurred boundaries and demonstrate reduced classification accuracy under complex stratigraphic conditions. This study proposes a Stratigraphically Constrained Continuous Clustering (SCCC) framework to address these limitations. The framework incorporates sedimentological principles of lateral continuity through a dynamic density-threshold hierarchical clustering algorithm that optimizes lithological unit boundaries using adjacency-based cluster merging criteria. A patch-level feature aggregation module, integrated within the proposed SCCC framework, constructs a multimodal feature space by aggregating geometric covariance matrices and spectral distribution entropy into compact patch-level feature vectors. Random forest classifier subsequently performs lithology discrimination. Experimental validation using the Qingshuihe Formation outcrop dataset demonstrates that SCCC achieves overall accuracy of 94.64%, F1-score of 94.58%, and mean intersection over union of 90.87%. These results surpass traditional machine learning (SVM, XGBoost) and deep learning methods (PointNet) by 26.22–68.36%, indicating substantial improvements in classification accuracy and boundary delineation within transition zones. SCCC particularly enhances recognition capabilities for sandstone-mudstone thin interbeds and conglomerate-sandstone transitional zones. Ablation experiments confirm that stratigraphic constraints effectively suppress noise while improving computational efficiency, reducing memory usage by 83.3% and processing time by 85.7%. This method provides a high-precision, interpretable technical pathway for intelligent geological exploration through deep integration of geological principles with computational models.