Intelligent RQD cataloging of rock masses based on a lightweight vision foundation model

To address the issues of inconsistent evaluation criteria in manual core cataloging of massive boreholes and delayed digital delivery, as well as the susceptibility of existing vision methods to interference in complex geological conditions, this study conducts research on an intelligent rock quality designation (RQD) cataloging method based on a lightweight adaptation of segment anything model (SAM). First, a two-stage distillation strategy is employed to substantially reduce the number of model parameters while effectively preserving the strong feature representation capability of the SAM encoder. On this basis, a lightweight feature aggregator and an efficient serial decoder are designed to construct a rock core instance segmentation model, termed LW-CoreSAM. Furthermore, by integrating a rock core length quantification method, an intelligent RQD cataloging framework is developed for on-site applications. Validations from multiple engineering case studies demonstrate that the LW-CoreSAM, with only 8.02 million parameters, achieves a mean AP of 81.16% with a standard deviation of only 4.14%, outperforming mainstream instance segmentation models. The output RQD values show high consistency with manual logging (R² > 0.95), demonstrating an effective balance between high accuracy and real-time performance. This study overcomes the computational bottleneck of SAM's end-side deployment and effectively eliminates subjective differences in evaluation criteria for core cataloging, providing technical support for the standardization and digitalization construction of geological survey data.