Rockfall Isolation Technique Based on DC-DBSCAN with k-Means Clustering and k-Nearest Neighbors Algorithm

Abstract

Recently, spatial-clustered point clouds have been applied to various applications, such as glacier movement and rockfall detection, which are crucial for ensuring human safety. The density-based spatial clustering of applications with noise (DBSCAN) is a well-known spatial clustering algorithm. It is effective but requires two predefined parameters needed to be appropriately set. The suitable values of these parameters depend on the distribution of the input point cloud. Thus, to address this issue, we previously proposed a non-parametric DBSCAN based on a recursive approach and called it divide-and-conquer-based DBSCAN or DC-DBSCAN. Even though it outperformed the traditional DBSCAN, the performance of the previous DC-DBSCAN or DBSCAN is limited when two groups or two clusters are too close. Therefore, this study proposes an improved version of DC-DBSCAN that utilizes the k-means clustering algorithm to further cluster some groups resulting from DC-DBSCAN. To determine which groups are to be clustered further, a k-nearest neighbors algorithm is used. The experimental results demonstrate that the proposed method enhances the impurity and normalized mutual information (NMI) scores compared with DBSCAN and DC-DBSCAN. The purity score of the proposed method is $97.91\%$, and the NMI score is $96.48\%$. Compared to DC-DBSCAN, our proposed method achieves a $12.37\%$ improvement in purity and a $3.61\%$ improvement in NMI. Also, it can spatially cluster some groups that DBSCAN and DC-DBSCAN cannot do.