Automated Sediment Characterization to Understand Long-Term Coastal Change in Response to the Elwha River Dam Removal

Abstract

Sediment characteristics – relative proportions of sediment fractions and the prevalence ofwood or other organic debris – are a fundamental property of coastal systems, influencing storm response, habitat quality, and long-term shoreline evolution. Sediment characteristics and coastal morphology vary in time and space with sediment supply, wave activity, and other local dynamics. Despite their hypothesized importance, the co-evolution of sediment characteristics and coastal morphology remains poorly understood in mixed sediment en- vironments due to challenges with data collection and analysis. The dataset used in this research is made up of 15,000 sediment images collected from 2009 to 2023 on the Elwha River delta and adjacent downdrift coastlines in Washington. This research focuses on the annual co-evolution of sediment characteristics and coastal morphology following a major flux of sediment to the system after the undamming of the Elwha River. SegFormer models were integrated into an automated image segmentation workflow and trained to identify and quantify the relative proportions of sediment fractions and organic debris across this long- term dataset. The results from these models show clear spatial and temporal variability in sediment composition across the Elwha River delta following dam removal, with shifts in the relative proportions of sand, gravel, and cobble over time. Following dam removal, sand in- creased near the river mouth and gradually propagated alongshore as the delta transitioned from a coarse-dominated to a sand-dominated shoreline. Sand accumulated on the upper beach of downdrift coastlines, suggesting that cross-shore processes were also important for sediment redistribution. A strong link between sediment delivery and coastal response was identified, demonstrating how machine learning-based image segmentation can be used to monitor mixed-sediment coastal change.