Predicting the index flood in ungauged UK catchments: on the link between data-transfer and spatial model error structure

This paper presents a revised procedure for predicting the index flood in ungauged UK catchments based on a revised data transfer methods. Using annual maximum peak flow data from 602 catchments in the UK, the performance of hydrological regression models linking the (log) index flood to a set of (log) catchment descriptors is investigated and compared to the benefit of enhancing the regression-based estimates with data transfer from a gauged site that is either i) geographically close, ii) considered hydrologically similar, or iii) both i) and ii). The study compares the performance of two regression models when combined with data transfer: i) a comprehensive model with four catchment descriptors, ii) a simple model using only catchment area as an explanatory variable. The results show that the simple regression model benefits more from additional data transfer than does the more comprehensive model. It is shown that, when data transfer is included and when the ungauged and the gauged site are extremely close, the two models perform equally well. In cases where the nearest gauged catchment is 1 to 20km away, data transfer allows the simpler model to provide predictions which are only moderately worse than those from the comprehensive model, whereas if data transfer is not used the simpler model is substantially worse than the comprehensive one. Little or no benefit was gained by selecting the gauged site by considering similarity of catchment area. Using data-transfer reduces the spatial correlation in the regression residuals. The results presented in this paper are believed to be of special interest to hydrologists working in regions where only a limited subset of catchments descriptor is routinely made available