Detection of Stage‐Discharge Rating Shifts Using Gaugings: A Recursive Segmentation Procedure Accounting for Observational and Model Uncertainties

The stage‐discharge rating curve is subject at many hydrometric stations to sudden changes (shifts) typically caused by morphogenic floods. We propose an original method for estimating shift times using the stage‐discharge observations, also known as gaugings. This method is based on a recursive segmentation procedure that accounts for both gaugings and rating curve uncertainties through a Bayesian framework. It starts with the estimation of a baseline rating curve using all available gaugings. Then it computes the residuals between the gaugings and this rating curve with uncertainties. It proceeds with the segmentation of the time series of residuals through a multi‐change point Bayesian estimation accounting for residuals uncertainties. Once the first set of shift times is identified, the same procedure is recursively applied to each sub‐period through a “top‐down” approach searching for all effective shifts. The proposed method is illustrated using the Ardèche River at Meyras in France (a typical hydrometric site subject to river bed degradation) and evaluated using several synthetic data sets for which the true shift times are known. The applications confirm the added value of the recursive segmentation compared with a “single‐pass” approach and highlight the importance of properly accounting for uncertainties in the segmented data. The recursive procedure effectively disentangles rating changes from observational and rating curve uncertainties. Plain Language Summary For many hydrological and hydraulic issues, such as flood forecasting, a reliable river discharge estimate is needed. In general, discharge is derived from the recorded water level (stage) through a stage‐discharge relation (rating curve). This relation is calibrated using direct observations (gaugings). Unfortunately, the rating curve is not only uncertain but it can also be subject to sudden changes or shifts due for example to intense floods that modify the river bed geometry. One solution to identify periods of rating curve stability is to apply a segmentation procedure to the gaugings. We propose in this paper an original recursive segmentation procedure that accounts for both gaugings and rating curve uncertainties. Key Points We propose a method for detecting rating shifts through the segmentation of residuals between the gaugings and a reference rating curve The method accounts for observational and rating curve uncertainties and expresses change points in terms of time (rather th