A Comprehensive Deep Learning and Machine Learning Framework for Shoreline Prediction: Enhancing Accuracy and Interpretability for Dublin Bay, Ireland
Year: 2025
Type: dissertation
Abstract: Predictive shoreline forecasting is essential for safeguarding coastal communities and preserving dynamic shore ecosystems, which are subject to ongoing natural variation and environmental change. Traditional physics-based models and pure data-driven models lack the capability to accurately represent the rapid and nonlinear changes due to increased human influence and evolving environmental conditions. This study presents an innovative hybrid GeoAI pipeline designed to improve shoreline forecasting, in Dublin Bay as the case study. The proposed methodology involves a three-stage process. First, spatial features from Sentinel-2 satellite images collected between 2017-2025 will be used to train a pre-trained Masked Autoencoder (MAE). Secondly, the trained MAE will then be combined with a Transformer Encoder to identify complex, non-linear time series trends in the data. Finally, the resulting feature vectors will be integrated into a Light Gradient Boosting Machine (LightGBM) meta-learner, along with conventional physical attributes (e.g., tidal cycles, wave activity, wind direction and speed and DEMs). A comparative analysis of the proposed hybrid model (Model C) showed that it delivered a notable improvement in performance, surpassing both traditional machine learning and deep learning models. Model C produced the largest amount of explained variance (R2 = 80.07%) and had the smallest average absolute error (MAE = 1.13 m), a 7.1 cm reduction in MAE and a 1.89% increase in R2 when compared to the classical model. The top features utilized by the hybrid model showed that deep learned features, generated by the Transformer Encoder, were significant contributors to the improved performance of the hybrid. In summary, the hybrid GeoAI framework introduced in this research has significantly improved both the predictive accuracy and interpretability of shoreline modelling. By combining spatial, temporal, and physical data, it addresses several weaknesses associated with previous methods and provides a robust tool for informing coastal management and resilience planning.
Source: NORMA
Author: Alleyne, Shamaie Maria
Language: English
Topic: Coastal and Marine Dynamics
Subfield: Earth-Surface Processes
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212095895