Author(s): Emma Mcallister

Linked Author(s): Emma Mcallister

Keywords: Remote Sensing; Coastal Engineering; Machine Learning

Abstract: The world's coastlines are under pressure from the threat of future Sea Level Rise (SLR). As coastal engineers, we need a way to monitor shoreline change over short and long-term timescales, for the current and future protection of coastal communities. Satellite imagery provides users with high temporal and spatial resolution data, which allows users to monitor the shoreline over both short and long-term time scales. Although we have access to high-resolution and frequent data, one of the main issues of using satellite imagery for coastal studies is the accurate extraction of the “true” shoreline position using a suitable coastal indicator. As the shoreline in a satellite image can be taken during any stage of the tidal cycle, providing users with an inconsistent shoreline position in every image, which then has to be further processed for tidal correction. This study looks at the capabilities of free high-resolution satellite imagery from Sentinel-2 which has been retrieved using Google Earth Engine (GEE), an online cloud-based platform used for geospatial analysis. The coastal indicator used to represent the shoreline position is the wet/dry boundary, the band parallel to the sea, which shows the extent of the previous high tide. This shore parallel band represents the wet sand, which can be delineated from images to give a consistent shoreline position. The shoreline has been extracted Classification and Regression Trees (CART) and an Artificial Neural Network (ANN). CART is a machine learning technique, which uses a decision tree algorithm to perform a classification by using a series of binary decisions, where the input is evaluated and one of two branches, are selected, to place the pixels of an image into different classes. ANNs consist of a network of neurons, which passes signals to each other. Neural networks are trained by initially assigning ‘random’ values for the weights, which are then adjusted when the data is backpropagated through the network which is based on the error associated with the output nodes. This training of the network is repeated until the error of the output is reduced to a minimum To date, preliminary results produced by the CART classifier at the study site Start Bay, show that the wet/dry boundary can be used as a proxy line for the extraction of the shoreline position. Results from the CART classifier will be compared to the ANN results to see which method produces the higher accuracy for the identification of the wet/dry boundary.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221201

Year: 2022