the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Marine Heatwaves in the Mediterranean Sea: A Convolutional Neural Network study for extreme event prediction
Abstract. In recent decades, the Mediterranean Sea has experienced a notable rise in the occurrence and intensity of extreme warm temperature events, referred to as Marine Heatwaves (MHWs). Hence, the ability to forecast Mediterranean MHWs in the short term is an area of ongoing research. Here, we introduce a novel machine learning (ML) approach, specifically tailored for short-term predictions of MHWs in the basin, using an Attention U-Net Convolutional Neural Network. Trained on daily Sea Surface Temperature anomalies and gridded fields of MHW presence and absence between 1982–2017, our model generates a spatiotemporal forecast of MHW occurrence up to 7 days in advance. To ensure robust performance, we explore various configurations, including different forecast horizons and U-net architectures, number of input days, features, and different subset splits of train-test datasets. Comparative analysis against a Persistence benchmark reveals that our model outperforms the benchmark across both forecast horizons. For the 7-day forecast, the model achieves a 15 % improvement in forecasting accuracy of MHW presence over the Persistence, while for the 3-day forecast, this improvement percentage drops to 4.5 %. Notably, the discrepancy between our model and the benchmark narrows for shorter horizons, as the Persistence method also achieves high accuracy in the 3-day forecast. Our proposed ML methodology offers a data-driven alternative for MHWs prediction with reduced computational requirements, which can be applied across different regions of the global ocean, providing relevant stakeholders and management authorities with essential lead time for implementing effective mitigation strategies.
- Preprint
(2157 KB) - Metadata XML
-
Supplement
(15 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-4003', Alexander Nickerson, 17 Feb 2025
The authors applied a relatively new kind of artificial intelligence to examine marine heat waves (MHWs) in the Mediterranean Sea. They highlight both the strengths (ability to issue reasonable forecasts) and weaknesses (e.g. computational intensity) of the method. There are some minor revisions and comments to address, but I recommend this manuscript for publication.
One significant addition, preferably in the paragraph starting on line 137, would be to note the important fact that MHW is defined based on a 30-year climatological mean. With most SST data sets now going back to the 1980s, the selection of a 30-year window can alter the results of an experiment like this, especially since many regions have statistically significant different temperatures now compared to 30-40 years ago. Which 30 year window was used for the identification of MHWs, or is it based on the window 1982-2017 like the SST anomalies?
It appears that the AI model used is fairly useful for short-term forecasts. Could you please note the merits of this relative to other AI methods used in climatology?
Line 30: Generally we use oC/decade or oC/century.
Line 50: "ML" needs to be defined here.
Line 87: "Artificial Neural Network" needs no capitalization
Fig. 1: Ierapetra Gyre is labeled with both an arrow and one of the red rings
Line 155: "SSTA" needs to be defined here.
Fig. 2: is the domain 128x272 the size of the data region in terms of data matrix size?
Eq. 2: y and γ are used interchangeably, as are a and αCitation: https://doi.org/10.5194/egusphere-2024-4003-RC1 -
RC2: 'Comment on egusphere-2024-4003', Anonymous Referee #2, 07 Mar 2025
I found your study very interesting and relevant to the growing field of AI-driven forecasting for extreme ocean events. Below are some comments that I believe could strengthen your manuscript:
- Comparison with ocean models: While you acknowledge ocean models in the introduction, there is no direct comparison between your AI-based approach and a traditional numerical ocean model for MHW forecasting. A quantitative comparison—whether in terms of forecast accuracy, computational efficiency, or ability to capture physical processes—would provide valuable context and help clarify how AI complements or improves upon traditional approaches.
- Resolution of the output: You mention that the final spatial resolution is downsampled to a 128x272 grid, but it would be helpful to explicitly state the corresponding resolution in kilometers. Additionally, how does the downsampling affect the model’s ability to capture finer-scale MHW dynamics, particularly in regions with complex topography such as the Adriatic or Aegean Seas?
- Uncertainty Quantification: Your study provides a robust evaluation of model performance, but there is limited discussion on uncertainty quantification. Given the stochastic nature of neural networks, have you assessed the sensitivity of your predictions to different training datasets, hyperparameter choices, or initial conditions? Methods such as ensemble modeling could provide insights into the confidence of the forecasts.
- Generalization Beyond the Mediterranean: You mention that the proposed method could be applied to other regions, but it would be helpful to discuss potential challenges in doing so. For example, would a model trained on Mediterranean SST anomalies generalize well to other basins with different oceanographic characteristics (e.g., stronger currents, different stratification, or more extreme variability)? A brief discussion of how the model could be adapted or retrained for different environments would be valuable.
Overall, this study presents a promising application of deep learning for MHW forecasting, and I appreciate the detailed methodology and validation process. Addressing these points could further enhance the robustness and impact of your work.
Citation: https://doi.org/10.5194/egusphere-2024-4003-RC2 -
EC1: 'Comment on egusphere-2024-4003', Yonggang Liu, 08 Mar 2025
Line 53, abbreviation “DL” should not be used, because it is used only 3 time in the manuscript.
Line 164, it seems that local winds have not been considered as an input variable. This may partially account for the relatively poor ML model performance in the coastal ocean areas, such as the Aegean Sea and the Adriatic Sea. It is mentioned later in the paper that air temperature has been tested as an input variable, but not much improvement in MHW forecast. Have you tested winds as an input?
Line 220, should “NN” be ‘CNN” o r a new acronym?
Line 534 – 540, another possible reason could be the excluding of winds as an input the ML model. This is because coastal ocean circulation is mostly driven by local winds. Also, ocean circulation may be influenced by offshore currents forcing at the locations of complicated topographic features. These may not be properly represented in the ML model. Thus, temperature changes (and MHW changes) become more complicated in shallow coastal oceans (e.g., Berthou et al., 2024; Liu et al., 2025). It would be good to include these in the discussion.
References:
Berthou, S., Renshaw, R., Smyth, T., Tinker, J., Grist, J. P., Wihsgott, J. U., et al.: Exceptional atmospheric conditions in June 2023 generated a northwest European marine heatwave which contributed to breaking land temperature records, Communications Earth & Environment, 5(1), 287. https://doi.org/10.1038/s43247-024-01413-8, 2024.
Liu, Y., Weisberg, R.H., Sorinas, L., Law, J.A., Nickerson, A.K.: Rapid intensification of Hurricane Ian in relation to anomalously warm subsurface water on the wide continental shelf, Geophysical Research Letters, 52, e2024GL113192, https://doi.org/10.1029/2024GL113192, 2025.
Citation: https://doi.org/10.5194/egusphere-2024-4003-EC1
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
144 | 56 | 6 | 206 | 18 | 7 | 8 |
- HTML: 144
- PDF: 56
- XML: 6
- Total: 206
- Supplement: 18
- BibTeX: 7
- EndNote: 8
Viewed (geographical distribution)
Country | # | Views | % |
---|---|---|---|
United States of America | 1 | 59 | 29 |
Greece | 2 | 47 | 23 |
China | 3 | 18 | 8 |
France | 4 | 17 | 8 |
Germany | 5 | 5 | 2 |
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1
- 59