CNN-Based Imagery Analysis for Monitoring Complex Coastal Processes

Abstract

Video monitoring is essential for a variety of coastal research. While monitoring can provide a wealth of information about coastal processes, extracting relevant information from in-situ beach photography remains challenging. This dissertation elucidates how the novel semantic segmentation method based on Convolutional Neural Networks (CNN) could analyze beach imagery and shows how these results can shed light on two crucial coastal processes that shape beach morphology: coastal flooding and beach recovery. In the first part of the investigation, we trained CNN to extract the water areas from each coastal image captured in Cedar Lakes, Texas, following the beach breach caused by Hurricane Harvey in 2017. We found that prediction accuracy for detecting water pixels was around 90%, proving that CNN-based image segmentation can effectively analyze short-range coastal images. It is also suggested that, with the aid of transfer learning, more than 100 training images were sufficient for training a model that provides accurate image segmentation prediction. Using the CNN, we measured the time series of water area fractions where we defined coastal flooding events. It was found that the size of coastal flooding follows an exponential distribution, whereas the inter-arrival times of flooding events may or may not follow an exponential distribution, depending on the time series scale. For the second part of the investigation, we applied several new techniques to better use high-resolution images and trained the new CNN model to track how the beach regions change over time. The model predicted images in the validation set with a mean precision of 95.1% and demonstrated that it could reliably detect sand composition changes in the images, albeit failing to predict blurry or rainy images. By applying this new model, we studied how different sand areas change over time during beach recovery. The sand pattern changed dramatically following a massive aeolian transport event, which implied the beach had converted from non-recovering to recovering states after the event. Furthermore, measurement of aeolian transport revealed that the frequency of aeolian transport is primarily determined by the wind velocity, the amount of dry sand covering the site, and the direction of the wind.