Separating the land from the sea: image segmentation in support of coastal hazards research and community early warning systems

Total water level (TWL) at the shoreline is the key metric for assessing coastal flooding and erosion. Predictions of TWL are necessary for long-term coastal planning and early warning systems including the USGS/NOAA Total Water Level and Coastal Change Forecast (TWL&CC). However, TWL is both difficult to predict and difficult to measure. Various TWL measurement techniques including cross-shore arrays of pressure sensors and wire gauges have been used in the past, but these techniques are costly, time consuming and unlikely to be scaled up to a national scale such as would be required for a comprehensive validation of the USGS TWL&CC forecast. Coastal imaging cameras were introduced as a lower cost and scalable solution to measuring TWL. However, extracting quantitative measurements of TWL from images has typically been done through hand-digitization. This has created a bottle-neck and there are now years of extremely valuable but relatively unprocessed imagery collected and achieved by the USGS and other institutions. The U.S. Geological Survey recently released time series images from five locations, which had previously been hand digitized. This dataset includes images from beaches on sandy shorefaces in the Gulf of Mexico (Madeira Beach, Florida (FL) and Sand Key, FL) and coral reef-fronted beaches in Puerto Rico (Tres Palmas, PR and Isla Verde, PR) and Hawai’i (Waiakāne, HI). These data were used to train and validate machine learning models. The final trained model was able to extract the TWL signal from the images with an accuracy equal to hand digitizers. With the new model images can be processed in ten seconds vs the ten minutes required for hand digitization. This makes the model appropriate for processing large quantities of archival data, as well as near-real time applications. With the ability to autonomously process TWL observations, coastal imaging systems can be converted into stand-alone early warning systems, assimilated into coastal hazards forecasts and used for continual model validation and improvement.