Regionalization of watersheds by finite mixture models

Cluster analysis methods are one of the most widely-used types of methods in regionalization of watersheds for regional flood frequency analysis. Several cluster analysis algorithms have been used in regional frequency analysis studies in the last decades, which most of them have been data-driven algorithms that identify clusters based on distances or dissimilarities between data points. Finite mixture models are a type of statistical models that are able to identify the clusters based on frequency distributions. In this study, the performance of finite mixture models for regionalization of watersheds was evaluated by applying Gaussian mixture models for regionalization of watersheds of Karun-e-bozorg in the southwest of Iran. Gaussian mixture models are the best-known type of finite mixture models. The performance of this method was evaluated according to the homogeneity of regions and the accuracy of flood quantile estimates. In addition, the results of the proposed method were compared with the results provided by using a well-known and efficient data-driven hybrid clustering algorithms. According to the results, the Gaussian mixture models clearly outperformed the hybrid clustering algorithm in identifying homogeneous regions. While one of the Gaussian mixture models assigned all the watersheds to homogeneous regions for five different choices of the number of regions, the hybrid clustering algorithm assigned all the watersheds homogeneous regions only for one choice of the number of regions. Also, the flood estimates provided based on the regions identified by the Gaussian mixture model and the hybrid clustering algorithm were compared with the at-site estimates. As evidenced by the results, the flood estimates related to the regions identified by the Gaussian mixture model are less deviated from at-site estimates in terms of the error measures. In general, results show that finite mixture models can be considered as an efficient option to perform regionalization of watersheds in regional flood frequency analysis studies.