In the past years, Judith’s research focus has been on the quantification of the quality of morphodynamic predictions. Her PhD thesis—which she defended on January 16, 2020—is about:
The behaviour of the widely used mean-squared-error skill score with the initial bed as the reference, which goes by the name Brier skill score.
The development of novel validation methods and corresponding error metrics that take the spatial structure of morphological patterns into account:
- A field deformation or warping method, which deforms the predictions as to minimize the misfit with observations;
- An optimal transport method, which moves misplaced sediment from the predicted to the observed morphology through an optimal, rotation-free sediment transport field;
- A scale-selective validation approach, which allows any metric to selectively address multiple spatial scales.
The PhD thesis contains the following findings:
- The use of a single performance metric leads to an inadequate interpretation of quality.
- A set of performance metrics for morphological models must include a metric—such as the root-mean-squared transport error (RMSTE)—that takes the spatial structure of morphological patterns into account.
- Optimizing the mean-squared error (MSE) or derived skill score (MSESS or BSS) of a morphological prediction leads to undesired underprediction of the variance of bed changes.
- The MSE-based skill score using the initial bed as the reference (a.k.a. the BSS) fails at making predictions comparable, whether across different prediction situations or across different times in a simulation.