Comparision of performance of multi criteria decision making ensemble-clustering algorithms in rainfall frequency analysis

Non availability of adequate extreme rainfall information at any place of interest are solved using regionalization where subjective grouping of similar attributes of nearby gauged stations is performed. K-Means and Fuzzy C-Means are commonly used methods in regionalization of rainfall, but application of genetic algorithm is very rarely explored. Genetic algorithms (GA) are highly efficient evolutionary algorithms, and through an appropriate objective function can effectively achieve the purpose of clustering. In the present study, Davies-Bouldin index is considered and validation is performed using a set of validation measures. Taking into account the varied output obtained in each validation measure, an ensembled approach involving multi criteria decision making is applied to obtain optimal ranked solutions, and the procedure is extended to K-Means and Fuzzy C-Means for comparision. From the results obtained, GA based clustering is found to outperform other two algorithms in formation of homogenous regions with better performances in leave-one-out cross validation (LOOCV) test and sensitivity analysis. Accuracy of regional growth curves of regions assessed using regional relative bias and RMSE suggest low uncertainty and accurate quantile estimates in GA regions. Further, information transfer index based on entropy evaluated among GA regions is found to be highest and K-Means lowest.