Comparative research on thunderstorms identification based on three clustering methods

In order to evaluate the recognition effect of different clustering algorithms on thunderstorm system, further improve the ability of lightning nowcasting, in this study, with the ground-based lightning location data and radar reflectivity data, three different clustering algorithms for thunderstorm identification are analyzed for a thunderstorm case in the region of 114°-117°E, 27°-30°N and during the time window of 19∶15 BT to 20∶57 BT on 21 September, 2018. The performance of these three algorithms, which are density-based spatial clustering of application with noise (DBSCAN), clustering by fast search and find of density peaks (CFSFDP) and extended clustering by fast search and find of density peaks (E_CFSFDP) are compared and discussed. Here are the results. (1) The DBSCAN algorithm has a higher accuracy of classification and identification when the distributions of lightning location show clear patterns and different clusters are apart from each other, but the accuracy will be low when the density of each lightning data cluster is uneven or the distances between different clusters are very large. (2) the CFSFDP algorithm works when each cluster in the data sets has one single density peak, otherwise, when there are no density peaks, CFSFDP will be failed. (3) The E_CFSFDP algorithm can solve the "no density peaks" issue of the CFSFDP algorithm, and the distribution of lightning location has little effect on the clustering method. Generally, the accuracy of the thunderstorm identification based on the E_CFSFDP algorithm is significantly higher than that of the DBSCAN and CFSFDP algorithms.