Beach monitoring: Do we need to survey to spring low tide?

When collecting coastal monitoring data, it is common practice to survey down to spring low tide to capture the maximum extent of the exposed subaerial beach. However, collecting topographic beach data is possible for only a few days per month. By reducing the seaward extent of the survey, the incurred costs and risks to the survey schedule could be greatly reduced. However, this would result in information loss at the lowest extremes of the subaerial beach. This study assesses the feasibility of predicting this part of the beach using deep learning neural networks based on partial beach profile data. A range of network architectures were tested alongside linear extrapolation, which was used as a baseline model. Each model was tested on three beaches with varying morphology, ranging from steep (reflective) to mildly sloping (dissipative). The presence of morphological features was found to play a dominant role in the accuracy of the predicted profiles; profiles with more pronounced cross-shore morphological features, such as sandbars, produced the highest error. While local connectivity of each network architecture was found to be the key factor in producing realistic profiles, the 1D Convolutional Neural Network was found to be the most effective with an average RMSE of between 0.026-0.119 m. This RMSE is not substantially larger than the vertical accuracy of current survey techniques (0.03 m), and the study found that errors of this magnitude have negligible effects when the survey data is used to calculate beach volumes and conduct numerical wave runup analysis to assess coastal flood risk. • Neural networks can reconstruct lower beach profile extents from partial data. • 1D CNN achieved a comparable accuracy to standard survey methods. • Enables optimisation of monitoring without compromising coastal hazard assessments.