Machine learning algorithms for automatic velocity picking: K-means vs. DBSCAN

Accurate and efficient algorithms for automatic velocity picking can significantly reduce the cycle time in seismic data processing. Recently, machine learning algorithms have been proposed to tackle the problem of automatic velocity picking. However, the proposed methods are based on the Kmeans clustering algorithm, which is known to suffer from several problems. Chief among them is the requirement that the user needs to input the number of clusters to separate the data points. This information is not easily available and therefore, the proposed velocity picking methods require several trial and error runs to select the optimal number of clusters. Contrary to that, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm does not require the user to input the number of clusters. Moreover, DBSCAN can separate clusters in a nonlinear fashion and is less susceptible to noisy data. Here, we compare the K-means and DBSCAN algorithms for velocity auto-picking. We find that DBSCAN yields superior performance in velocity picking compared to K-means. Furthermore, DBSCAN requires significantly less manual intervention as opposed to K-means. The difference in computational cost between the algorithm is negligible for the problem at hand. These findings can significantly reduce the cost of manual velocity picking, driving down the cost of processing large seismic datasets.