Marine Heatwave Analysis and Prediction Using Deep Learning: A Case Study Around Ireland

Marine heatwaves (MHWs)—prolonged periods of anomalously warm sea surface temperatures (SST)—pose significant ecological and economic challenges, particularly for aquaculture sectors sensitive to temperature variability around Ireland. This study integrates 43 years of historical daily SST data (1982–2024) from NOAA, ICES, and the Marine Institute to develop a comprehensive deep-learning framework for predicting SST and detecting MHWs in the Irish maritime region.

A comparative analysis of two MHW detection methodologies—Hobday et al. (2016) and Darmaraki et al. (2019)—was conducted, highlighting regional trends and spatial patterns of MHW characteristics like frequency, duration, and intensity. The Darmaraki method, with its 99th percentile threshold and flexible event merging criteria, was found to better capture localized and extreme temperature anomalies relevant to aquaculture, while the Hobday method identified a broader range of moderate events. The findings show that MHW frequency has increased significantly over time, particularly in the southeastern and northern waters, with some regions experiencing a doubling of annual MHW events as detected by the Darmaraki method. Long-duration MHWs, exceeding 60 days, are frequently observed along the western and southeastern coasts, demonstrating persistent thermal stress in these areas. The most intense MHWs, with temperature anomalies surpassing 2.5°C above climatological baselines, are concentrated in the southwestern and offshore regions. These areas emerge as critical hotspots, underlining the need for targeted monitoring and adaptive strategies for aquaculture management.

Deep learning models were introduced to predict SST and assess MHW risks to address the need for actionable forecasts. Long Short-Term Memory (LSTM) networks are particularly well-suited for analyzing time series data, as they effectively capture temporal dependencies and long-range patterns in sequential datasets. When coupled with the PyTorch framework, these models offer flexibility and scalability, making them ideal for large and complex SST datasets. Furthermore, combining LSTM with Convolutional Neural Networks (LSTM-CNN) enables the integration of both temporal and spatial features, which is crucial for understanding the intricate dynamics of MHWs.

The LSTM and LSTM-CNN frameworks demonstrated their effectiveness in forecasting SST across various temporal horizons, with predicted values evaluated against MHW criteria to identify potential events and their impacts. By leveraging these models, this study transitions from reactive to proactive MHW detection, providing early warnings and enabling aquaculture stakeholders to implement timely mitigation measures.

This interdisciplinary study bridges marine science and data engineering, combining observational data, machine learning, and robust detection frameworks to enhance the monitoring, forecasting, and management of extreme ocean events. The outcomes provide critical tools for sustainable aquaculture management and contribute to the broader understanding of climate impacts on marine environments.