A comparative analysis of clustering algorithms to identify the homogeneous rainfall gauge stations of Bangladesh

Dealing with individual rainfall station is time consuming as well as prone to more variation. It seems reasonable and advantageous to deal with a group of homogeneous stations rather than an individual station. Such groups can be identified using clustering algorithms, techniques used in the multivariate data analysis. Particularly, in this study, covering both hard and soft clustering approaches, three clustering algorithms namely Agglomerative hierarchical, K-means clustering and Fuzzy C-means methods are chosen due to their popularity. These algorithms are applied over precipitation data recorded by the Bangladesh Meteorology Department, and a comparison among the algorithms is made. Annual and seasonal precipitations from 1977 to 2012 recorded in 30 stations are used in this study. Optimal numbers of clusters in the four precipitation series are determined using the Gap statistic for K-means clustering and using the extended Gap statistic for Fuzzy C-means clustering, and are found as 3, 1, 3 and 2 for annual, pre-monsoon, monsoon and post-monsoon, respectively. This study investigates the clustering methods in terms of the similarity, members and homogeneity, among the clusters formed. The clusters are also characterized to see how they are distributed. Moreover, in terms of cluster homogeneity, Fuzzy C-means algorithm outperforms the other clustering methods.