Optimal depth-based regional frequency analysis

Abstract

Abstract. Classical methods of regional frequency analysis (RFA) of hydrological variables face two drawbacks: (1) the restriction to a particular region which can correspond to a loss of some information and (2) the definition of a region that generates a border effect. To reduce the impact of these drawbacks on regional modeling performance, an iterative method was proposed recently. The proposed method is based on the statistical notion of the depth function and a weight function φ. This depth-based RFA (DBRFA) approach was shown to be superior to traditional approaches in terms of flexibility, generality and performance. The main difficulty of the DBRFA approach is the optimal choice of the weight function φ (e.g. φ minimizing estimation errors). In order to avoid subjective choice and naïve selection procedures of φ, the aim of the present paper is to propose an algorithm-based procedure to optimize the DBRFA and automate the choice of φ according to objective performance criteria. This procedure is applied to estimate flood quantiles in three different regions in North America. One of the findings from the application is that the optimal weight function depends on the considered region and can also quantify the region homogeneity. By comparing the DBRFA to the canonical correlation analysis (CCA) method, results show that the DBRFA approach leads to better performances both in terms of relative bias and mean square error.