Models and scales for cross‐shore sandbar migration

Abstract

We investigate the long‐term (months to years) predictability of cross‐shore sandbar migration with two models that operate on different abstraction levels: (1) a coupled, cross‐shore waves‐currents‐bathymetric evolution model and (2) two data‐driven neural network models, based on simplified cross‐shore profile representations and daily averaged wave properties. For model calibration, training, and validation, we use a high‐resolution 15 yearlong profile data set collected at the Hasaki Oceanographic Research Station in Japan. Sandbar behavior at this field site is characterized by cycles of net‐offshore migration with a duration of 1–4 years. We find that all models can produce several general features of sandbar behavior at the studied field site, such as rapid offshore migration, slower onshore return, and net‐offshore migration. However, it is difficult to quantitatively predict the offshore‐directed trends in sandbar location over time scales of months to years. While simple linear models outperform more detailed nonlinear models, for all models it is difficult to predict long‐term sandbar behavior, because of error accumulation in the model's processes over time. Representing processes on a more abstract level (scale aggregation) alleviates error accumulation but does not completely overcome this problem.