Input Variable Selection for Median Flood Regionalisation

Flood estimation for ungauged catchments is a challenging task for hydrologists. A modern geographical information system is able to extract a large number of catchment characteristics as input variables for regionalization analysis. Effective and efficient selection of the best input variables is urgently needed in this field. This paper explores a new methodology for selecting the best input variable combination on the basis of the gamma test and leave-one-out cross validation (LOOCV) to estimate the median annual maximum flow (as an index flood). Since the gamma test is capable of efficiently calculating the output variance on the basis of the input without the need to select a model structure type, more effective regionalization models could be developed because there is no need to define an a priori model structure. A case study from 20 catchments in southwest England has been used to illustrate and validate the proposed scheme. It has been found that the gamma test is able to narrow down the search options to be further explored by the LOOCV. The best formula from this approach outperforms the conventional approaches based on cross validation, data filtering with Spearman's rank correlation matrix, and corrected Akaike information criterion. In addition, the developed formula is significantly more accurate than the existing equation used in the Flood Estimation Handbook.