A Review of Cluster Filtering Methods for Point-Cloud-Based Rockfall Monitoring

ABSTRACT: The standard methodology for through-time rockfall analysis from slope point clouds consists of computing changes between lidar scans, followed by grouping areas of significant change using unsupervised clustering (e.g., DBSCAN). Clusters found are indicative of a potential rockfall and can be used for the construction of a comprehensive rockfall database. However, clustered change not representing true rockfalls is common, often accounting for the majority of identified objects. Identifying these erroneous clusters requires hours of manual verification; therefore, the search for an effective filtering method to reduce the need for manual verification is a highly relevant topic in the field of point-cloud-based slope monitoring research. This paper presents a review of recent literature on the filtering of rockfall databases to remove erroneous clusters and maintain an overall high quality rockfall database. The review has highlighted the lack of standards for filtering these clusters and a corresponding high degree of variability in filtering methods. 1. BACKGROUND The risk of rockfall from slopes along transportation corridors can be characterized by both the magnitude (volume) and frequency of block detachments. Rockfall inventories are commonly used to define a magnitude-frequency curve for a given slope of interest. Using a power-law fit to such a curve, the frequency of certain sized rockfall events can be estimated (Dussauge et al., 2003; Dussauge-Peisser et al., 2002; Hungr et al., 1999). Using terrestrial laser scanning (TLS), high-resolution point cloud data is now available for regularly monitored slopes, allowing comprehensive rockfall databases to be created. To characterize these slopes, the construction of rockfall databases typically follows a standard methodology (Abellán et al., 2010; Brodu & Lague, 2012; Schovanec et al., 2021; Tonini & Abellan, 2014; Weidner et al., 2019): (i). Collection of multiple epochs of TLS. (ii). Alignment of collected point clouds. (iii). Calculation of change between subsequent scans to determine areas of apparent volume loss. (iv). Clustering detected areas of change using a clustering algorithm (commonly Density Based Spatial Clustering of Applications with Noise – DBSCAN) (Ester et al., 1996). (v). Classification and filtering of detected clusters, for example to discern rockfall clusters from non-rockfall clusters (vegetation, shadows, occlusion, snowfall, etc.). Through the application of this methodology, a rockfall database can be created for a monitored slope. While the application of either manual or automated cluster filtering (step "v") can be critical to avoid inclusion of erroneous clusters in a rockfall database, the acceptance of all clusters (i.e. no filtering) is a common approach in practice (see Table 1 for a summary of approaches adopted in the literature). This paper presents a review of the current approaches and challenges associated with rockfall cluster filtering, as well as its practical importance.