Predicting Shoreline Changes Along the California Coast Using Deep Learning Applied to Satellite Observations

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Adusumilli, Susheel, Cirrito, Nicholas, Engeman, Laura, Fiedler, Julia W., Guza, R. T., Lange, Athina M. Z., Merrifield, Mark A., O'Reilly, William, Young, Adam P.
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000172

Understanding and predicting changes in shoreline location are critical for coastal planners. In situ monitoring is accurate but not widely available. Satellite observations of shorelines have global coverage, but their accuracy and predictive capacity have not been fully explored. Abundant beach surveys and extensive wave observations in Southern California provide a unique ground truth for the interpretation of satellite‐derived recently wetted waterlines. We combine 23 years of waterline position estimates from satellite imagery with nearshore wave hindcasts and tides to train and test a deep neural network (DNN). The trained DNN uses only tides and waves as predictors at transects with satellite coverage and wave estimates to predict beach width and, for the first time, seasonal average beach slopes. Beach width changes hindcast using DNN have at least fair skill (>0.3) for 50% of transects, where the skill was calculated relative to the mean of extensive new in situ survey data from in San Diego County. The DNN also predicted shoreline changes to within 10 m (the nominal uncertainty in satellite‐derived shorelines) for 64% of transects lacking ground truth. DNN and survey estimates of seasonal beach slope changes also agree qualitatively. Plain Language Summary Monitoring changes in the width and slope of sandy beaches is crucial for effective coastal management and planning to mitigate risks of coastal erosion and flooding. Survey observations of beach widths using mobile laser scanners or drones are accurate but expensive to collect and unavailable for a large fraction of California's coastline. In this study, we combine satellite observations of changes in shoreline positions and a regional wave model to relate (using machine learning) waves and changes in beach width and slope. We use extensive survey observations in Southern California to validate the machine learning framework. We find that our machine learning framework allows us to make useful predictions of beach width change using only widely available tide and wave model outputs as input for 50% of the beaches considered. We also quantify seasonal changes in beach slope for the first time, important because steeper slopes in winter (combined with large waves) lead to increased flooding. Key Points We apply deep learning to satellite‐derived shorelines to hindcast changes in beach width using tide and wave model outputs as predictors We validate these hindcasts against a large volume of unique in situ observations collected over Southern California beaches The deep learning algorithm provides the first satellite‐derived estimates of seasonal changes in beach slope, key for flood predictions