Comparison of automatic procedures for selecting flood peaks over threshold based on goodness‐of‐fit tests

In comparison with the traditional analysis of annual maximums, the peaks over threshold method provides many advantages when performing flood frequency analysis and trend analysis. However, the choice of the threshold remains an important question without definite answers and common visual diagnostic tools are difficult to reproduce on a large scale. This study investigates the behaviour of some automatic methods for threshold selection based on the generalized Pareto model for flood peak exceedances of the threshold and the Anderson–Darling test for fitting this model. In particular, the choice of a critical significance level to define an interval of acceptable values is addressed. First, automatic methods are investigated using a simulation study to assess fitting and prediction performance in a controlled environment. It is shown that P values approximated by an existing table of critical values can speed up computation without affecting the quality of the outcomes. Second, a case study compares automatically and manually selected thresholds for 285 sites across Canada by flood regime and super regions based on site characteristics. Correspondences are examined in terms of prediction of flood quantiles and trend analysis. Results show that trend detection is sensitive to the threshold selection method when studying the evolution of the number of peaks per year. Finally, a hybrid method is developed to combine automatic methods and is calibrated on the basis of super regions. The outcomes of the hybrid method are shown to more closely reproduce the estimates of the manually selected thresholds while reducing the model uncertainty.