In-situ block intelligent identification: Large-scale artificial intelligence (AI) models enable efficient geological surveys via image segmentation

In-situ block identification is important for characterizing the degree of rock mass jointing and for predicting rockfall volumes on engineering sites. Presently, research primarily focuses on using remote sensing techniques, such as drone-based photogrammetry, generating large-scale point cloud data of rock exposures as a substitute for dangerous and labor-intensive field surveys, while overlooking the huge difficulty of extracting block information from this data. Specifically, the difficulty lies in the nonintuitive characteristics of point cloud data and the lack of an efficient extraction technology. Large-scale models (LSMs) provide powerful visual perception capabilities across complex environments in computer vision, opening a relevant question for geological surveys: Can LSMs enable efficient identification of in-situ blocks? This study explores such possibility and proposes a novel approach, called Segment Anything for Three-Dimensional Blocks (SA4B-3D), that (i) converts (complex) 3D block identification challenges into (simpler) two-dimensional (2D) image segmentation problems, and (ii) leverages the Segment Anything Model (SAM) to intelligently identify individual blocks within point clouds of rock exposures. This paper evaluates the effectiveness of the method using a cardboard experiment, explores a real rock mass exposure and discusses the applicability of the proposed method, with results showing that the proposed method can obtain 7303 block observations with a total volume of 3462.7739 m³, tasks that were previously unachievable. Moreover, this LSM-based method paves the way for hour-level accurate geological surveys and advances engineering geology practice by reducing on-site labor and enhancing safety in high-risk projects.