Neural-network predictability experiments for nearshore sandbar migration

Abstract

Cross-shore migratory behavior of nearshore sandbars is commonly studied with nearshore bathymetric-evolution models that represent underlying processes of hydrodynamics and sediment transport. These models, however, struggle to reproduce natural cross-shore sandbar behavior on timescales of a few days to weeks and have uncertain skill on longer scales of months to years. One particular concern for the use of models on prediction timescales that far exceed the timescale of the modeled processes is the exponential accumulation of errors in the nonlinear model equations. The relation between cross-shore sandbar migration, sandbar location and wave height has previously been demonstrated to be weakly nonlinear on timescales of several days, but it is unknown how this nonlinearity affects the predictability of long-term (months to years) cross-shore sandbar behavior. Here we study the role of nonlinearity in the predictability of sandbar behavior on timescales of a few days to several months with data-driven neural network models. Our analyses are based on over 5600 daily-observed cross-shore sandbar locations and daily-averaged wave forcings from the Gold Coast, Australia, and Hasaki, Japan. We find that neural network models are able to hindcast many aspects of cross-shore sandbar behavior, such as rapid offshore migration during storms, slower onshore return during quiet periods, seasonal cycles and annual to interannual offshore-directed trends. Although the relation between sandbar migration, sandbar location and wave height is nonlinear, sandbar behavior can be hindcasted accurately over the entire lifespan of the sandbars at the Gold Coast. Contrastingly, it is difficult to hindcast the long-term offshore-directed trends in sandbar behavior at Hasaki because of exponential accumulation of errors over time. Our results further reveal that during periods with low-wave conditions it becomes increasingly difficult to predict sandbar locations, while during high waves predictions become increasingly accurate.