A convolutional, scale-adaptive framework for large-scale shoreline change modeling

Mohsen Taherkhani , Sean Vitousek, Peter Ruggiero

Understanding sandy beach evolution under geologic, hydrodynamic, morphologic, and anthropogenic influences, and particularly future climate variability and sea level rise, has become increasingly important. Several numerical models have been developed to simulate sandy beach evolution, ranging from high-fidelity, process-based models to reduced-complexity approaches, to data-driven and hybrid methods, and spanning scales from relatively small (hundreds of m-days) to relatively large (hundreds of km-decades). Data-driven approaches, however, often lack explicit representation of the physical processes governing shoreline change, a key sandy beach evolution metric. Here, we present a hybrid, scale-adaptive convolutional framework for shoreline change modeling that effectively bridges reduced-complexity and data-driven approaches and apply it to the U.S. Pacific Northwest sandy beaches (~750 km, Oregon and Washington). Using 40 years (1984–2024) of primarily satellite-derived shoreline data, we hindcast shoreline position evolution across 22 littoral cells comprising over 10,000 transects at 50-m alongshore resolution. Hindcast results yield a median validation Root Mean Square Error of 14.1 m and a Mielke's index of 0.51, similar to recent shoreline-change benchmarks. Our framework is based on a sum of convolution operations over hydrodynamic forcing with optimized kernel functions to model shoreline response. The resulting kernel shapes quantify the beach's morphodynamic memory, revealing shoreline response timescales and magnitudes that vary systematically with geographical latitude and coastal orientation across the region. The framework's computational efficiency and physical interpretability make it well-suited for future stochastic shoreline projections under changing climatic conditions, ultimately supporting informed coastal management and adaptation policies.