Wave runup and total water level observations from time series imagery at several sites with varying nearshore morphologies

Coastal imaging systems have been developed to measure wave runup and total water level (TWL) at the shoreline, which is a key metric for assessing coastal flooding and erosion. However, extracting quantitative measurements from coastal images has typically been done through the laborious task of hand-digitization of wave runup timestacks. Timestacks are images created by sampling a cross-shore array of pixels from an image through time as waves propagate towards and run up a beach. We utilize over 7000 hand-digitized timestacks from six diverse locations to train and validate machine learning models to automate the process of TWL extraction. Using these data, we evaluate two deep learning model architectures for the task of runup detection. One is based on a fully convolutional architecture trained from scratch, and the other is a transformer-based architecture trained using transfer learning. The deep learning models provide a probability of each pixel being either wet or dry. When contoured at the 50% level (equal chance of being wet or dry), the deep learning models more accurately identified TWL maxima than minima at all sites. This resulted in accurate predictions of 2% exceedance runup, but under predictions of significant swash and over predictions of wave setup. Improved agreement with the complete TWL time series was obtained through post-processing by utilizing the wet/dry probability of each pixel to weight the contouring toward lower dryness probabilities for runup minima (maxima agreed well with observations without tuning). Overall, a transformer-based model using transfer learning provided the best agreement with wave runup statistics, including a) the 2% exceedance runup, b) significant swash, and c) wave setup at the shoreline. For a random subset of images, the model was found to be within the uncertainty range of hand-digitization. The relative success of the transfer learning model suggests that finetuning a large model has advantages compared to training a smaller model from scratch. Models provide per- pixel probabilistic estimates in less than 10 s per timestack on a single computational unit, versus the more than 5 min required for hand-digitization. The model is therefore well-suited for near real-time applications, allowing for the development of early warning systems for difficult to forecast events. Real-time wave runup and total water level observations can also be incorporated into coastal hazards forecasts for data assimilation and continual model validation and improvement.