Harnessing artificial neural networks for coastal erosion prediction: A systematic review

Abstract

Artificial Neural Networks (ANNs) have proven highly effective for predicting coastal erosion, surpassing traditional models in capturing complex nonlinear relationships. This systematic review, conducted using the PRISMA protocol, evaluates 40 coastal related studies to assess ANN architectures, input variables, training techniques, and performance metrics. Findings indicate that Multi-Layer Perceptron (MLP) remains the most widely used ANN architecture, while hybrid approaches, including genetic programming and two-step networks, enhance prediction accuracy. Although Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs) have been explored, their applications remain limited. Research is predominantly concentrated in Asia and Europe, underscoring the need for expansion to Africa and South America. Despite advancements, challenges persist, including data scarcity, optimal data combinations, and model interpretability. Most studies focus on short-term predictions, often neglecting long-term coastal changes driven by climate change and sea-level rise. Additionally, ANN performance in predicting storm-induced erosion remains inconsistent, as extreme storm events introduce rapid, nonlinear changes that are difficult to model. Key research gaps include the integration of real-time data sources (e.g., wave, sediment, shoreline profiles, and storm data), improved model transparency, and better consideration of long-term shoreline evolution. Addressing these challenges will enhance ANN-based coastal prediction models, supporting adaptive management, early warning systems, and sustainable erosion mitigation strategies.