Regional Flood Frequency Analysis Using Entropy-Based Clustering Approach

Abstract

Hydrologists widely use regional flood frequency analysis to predict flood quantiles for ungauged and sparsely gauged target locations. The analysis involves: (1) use of a regionalization approach to identify a group of watersheds (region) resembling the watershed of target location by searching in watershed-related attribute space, and (2) use of information pooled from the region to perform regional frequency analysis (RFA) for flood quantile estimation. Conventional regionalization approaches prove ineffective for identification of regions in situations where there are outliers in the attribute space. This paper proposes an entropy-based clustering approach (EBCA) to identify regions by accounting for outliers and recommends applying a recently proposed RFA approach on the regions for quantile estimation. The EBCA yielded seven new homogeneous regions in four major river basins (Mahanadi, Godavari, Krishna, and Cauvery) of India. The regions are shown to be effective compared to six existing regions (used by an Indian government organization) and those delineated using global K-means clustering and region-of-influence approaches, in terms of regional homogeneity and utility in arriving at reliable flood quantile estimates for ungauged sites.