Application of machine learning techniques and high-resolution satellite imagery for the extraction of shoreline indicators and the prediction of future shorelines

Abstract

Coastal communities face increased risk from the challenges of erosion and flooding, combined by the rising threat of sea level rise (SLR). To understand the dynamics along global coastlines and accurately measure the rate of change, it is important to identify shorelines on a large scale over long periods (years - decades). The increase of remote sensing data, notably Multispectral Satellite Imagery (MSI), has provided opportunities to monitor shoreline change over long temporal periods. This thesis investigates the application of open-source MSI and Machine Learning (ML) techniques to quantify shoreline change across various beach types, through the automated extraction of the wet/dry boundary.

In this thesis, two key hypotheses are investigated. Firstly, the influence of coastline geomorphology and satellite imagery resolution on the feasibility of Shoreline Indicator (SI) identification is explored. It is expected that coastlines characterised by distinctive geomorphic features, combined with higher-resolution satellite imagery, will enhance the accuracy of visually discernible SI identification and extraction processes. Secondly, the hypothesis is posed that SIs can be reliably identified through ML methodologies. It is anticipated that by training ML algorithms on a diverse dataset of MSI, it will be possible to develop models that can accurately detect and classify SIs across different coastal regions worldwide.

This thesis initially assesses how previous Earth Observation (EO) studies mapped Shoreline Indicators from MSI. Recent emerging processing technologies, such as ML, are examined and considered regarding their capacity to identify and extract various shoreline features from local to global scales.

Two ML techniques, an Artificial Neural Network (ANN) and Classification and Regression Trees (CART) are developed for a two-step classification approach to identify the wet/dry boundary; these are tested against validation data derived from high resolution LiDAR and drone imagery. The ANN derived shorelines are then used to detect the wet/dry boundary line to allow for seasonal and multi-year analysis at three locations in the UK: Start Bay, Aldbrough and Sizewell.

This thesis subsequently investigates future prediction algorithms for forecasting shoreline change. A review of methods that are used for forecasting shoreline change is carried out and four ML techniques are tested for the prediction of future shorelines based on satellite derived shorelines. This investigation examines the potential of ML for predicting future shoreline change from MSI.

In this thesis, a significant achievement has been demonstrated, showcasing the effective utilisation of an ANN to accurately detect and extract the wet/dry boundary SI from Sentinel-2 MSI at Start Bay (±2.15m), Aldbrough (±10.67m), and Sizewell (±3.56m). These measurements align well with the 10m pixel value of Sentinel-2, underscoring the reliability and applicability of the ANN approach. This thesis highlights the adaptability and effectiveness of the developed ANN model in tackling diverse shoreline types. Additionally, the analysis of these shorelines proved invaluable in understanding shoreline dynamics over both seasonal and multiyear periods, revealing short-term erosion trends and long-term evolution patterns. Complex behavioural patterns and subtle trends of change were successfully replicated, contributing to a deeper understanding of coastal dynamics at the sites. Furthermore, this research has highlighted the potential of utilising satellite-derived shorelines for future shoreline prediction applications. The machine learning models employed in this study have demonstrated their potential to discern intricate change patterns, and the wealth of available satellite imagery data presents an opportunity to construct more robust and accurate predictive models. This not only enhances our capacity to anticipate shoreline changes but also paves the way for more informed coastal management strategies.