the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Jan 2026
An interpretable machine learning for marine heatwave prediction for the south China sea
Abstract. A primary challenge of machine learning to predict marine heatwave (MHW) for the south China sea (SCS) is the limited availability of observational data for model training. To address this issue, this study explores the viability of leveraging multi-member ensemble simulations from the Coupled Model Intercomparison Project Phase 6 (CMIP6), to construct an extensive, physically consistent training dataset for various machine learning models. After training on multiple CMIP6 ensemble members, the constructed models are evaluated for their predictive capacity regarding MHW in the SCS. The results also show that these machine learning-based methods can perform comparably to the existing dynamic models in terms of prediction performance, and in some cases even outperform the latter. Furthermore, by incorporating machine learning interpretability techniques, the key physical processes can also be elucidated from these predictions. That is to say, the new method is not a traditional "black box", but rather an effective tool that can possess certain physical transparency and scientific interpretability.
- Preprint
(4378 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 21 Aug 2026)
-
RC1: 'Comment on egusphere-2026-262', Anonymous Referee #1, 27 Feb 2026
reply
-
AC1: 'Reply on RC1', Peihao Yang, 14 Mar 2026
reply
Thank the editor and reviewers for their handling and review. Regarding the questions raised by the reviewers, our responses are as follows:
1, Firstly, short-term forecasting and seasonal/multi-year are fundamentally different. Generally, we cannot and do not need to predict the exact state of the climate 730 days in advance. I would suggest the authors study carefully the conventions of forecasting at different timescales, and to consider the target timescale most suited to their method.
Reply: Thank you for your comments. We agree that predictive capability and physical predictability differ substantially across time scales. Experiments with longer lead times (e.g., 90–730 days) are not intended to provide practical predictions of the exact future state. Instead, they are designed to illustrate how the predictive signal learned from CMIP6 gradually decays as the forecast horizon increases, thereby indicating the upper limit of the model’s usable predictability range.
2, Do the predictions presented represent an area-averaged time series? The SCS is a large area with many studies showing distinct MHW dynamics in the area. I am concerned that an area-averaged time series is not truly representative of individual events, nor relevant to impacts. Such an approach should at least cover sub-domains, if not individual grid cells on a common grid.
Reply: Thank you for your comments. In our study, the predictions are not based on an area-averaged time series. Instead, each grid cell within the South China Sea is treated as an independent prediction target, and the model uses the corresponding local predictor series for point-wise prediction. To avoid this confusion, we have revised the Section 3.2 Method and clarified this point in current version.
3, The forecast validation (Figure 2 and the corresponding analysis) is very unclear. This is unfortunate because it is the foundation of the paper. What does it mean to have a prediction accuracy of CMIP6 models or of ML models alone? Also, ERA5 is not used as a prediction.
Reply: Thank you for your comments. It should be clarified that ERA5 was not used as a prediction model. In the revised manuscript it has been corrected to “ECMWF forecast.” In addition, the definition of the accuracy has been clarified, indicating that the results from CMIP6 and related datasets represent event-agreement under a unified labeling framework rather than prediction accuracy in the machine-learning sense.
4, Accuracy is not defined.
Reply: Thank you for your comments. The accuracy used in this study represents the consistency of daily event identification. Specifically, at each time step, the MHW state predicted by the model is compared with the observed record to determine whether they are consistent, and the results are then averaged over the entire time series. This metric reflects the degree to which different methods consistently identify the occurrence or absence of events on a daily basis, allowing for a fair comparison under the same label construction scheme and the same class imbalance conditions. It should be noted that this metric is only used to evaluate pointwise classification consistency and does not replace the probabilistic or intensity-based error measures commonly used in extreme event studies.
5, A 7-day moving average is applied to target data. This removes variability that would be crucial for short-term forecasts, which are within the scope of this study. Please check the sensitivity of this approach.
Reply: Thank you for your comments. A typical MHW event is defined by a persistence of at least five consecutive days, meaning that its identification primarily relies on sustained warm anomalies rather than day-to-day fluctuations. The moving average suppresses high-frequency variability, making the labels more consistent with the persistence characteristic of MHW events and reducing spurious detections caused by transient noise. Because the definition of MHW itself is based on persistence, this processing step only stabilizes the labels and does not alter the event identification logic or affect the underlying climatic and physical interpretation.
6, It is not clear why Random Forest is considered the best performing model and subsequently used for the explainability analysis.
Reply: Thank you for your comments. Machine learning method shows predictive performance, where RF demonstrates stable performance, with relatively small error fluctuations and higher prediction consistency. In addition, tree-based models possess inherent advantages in structural transparency, variable interpretability, and quantitative evaluation of feature contributions. RF relies on explicit tree-splitting structures and node impurity reduction metrics, which enable it to provide robust and consistent rankings of feature importance and make it highly compatible with tree-based SHAP approaches. Therefore, this study selects RF to interpret the influence mechanisms of key meteorological and oceanic factors, and selects three representative forecast lead time—7-day, 30-day, and 90-day—for in-depth interpretation.
7, Line 202-204: Provide References.
Reply: Thank you for your comments. We have added References to this Section.
1)Wang, Z., Han, L., Ding, R., and Li, J.: Evaluation of the performance of CMIP5 and CMIP6 models in simulating the South Pacific Quadrupole–ENSO relationship, Atmospheric and Oceanic Science Letters, 14, 100057, 2021.
2)Yang, X. and Huang, P.: Improvements in the relationship between tropical precipitation and sea surface temperature from CMIP5 to CMIP6, Climate Dynamics, 60, 3319–3337, 2023.
Citation: https://doi.org/10.5194/egusphere-2026-262-AC1
-
AC1: 'Reply on RC1', Peihao Yang, 14 Mar 2026
reply
-
RC2: 'Comment on egusphere-2026-262', Anonymous Referee #2, 29 Jul 2026
reply
Review of "An interpretable machine learning for marine heatwave prediction for the south China sea" by Yang and Ye.
This work presents a comparison of several machine learning approaches to predict marine heatwaves in the South China Sea. This paper lacks a proper explanation of its methodology, and has an insufficient demonstration of results and their validation., so I cannot recommend it for publication.Here are some major comments about its contents:
First, there were several comments that were already raised by the other reviewer, (which I also had noted down) that I will not repeat but which are important (accuracy is not defined, use of ERA-5 as a forecasting system, which is not, are the results an area average...). But here are other comments that also need clarification:
- The paper does not mention how the datasets used are split into training, validation and testing datasets, and this is important to know how the model's performance has been assessed. The paper mentions validation and testing datasets but they are not described.
- The methodology is not clear, more details about each of the architectures used have to be provided. Also, what is the length of the study? What do the authors mean in line 210 "After training the machine learning models using CMIP6 multi-model data, the offline trained models are applied to observational data for prediction testing." And in line 242 "EOF analysis and lag analysis are applied to the predictors, and the derived variables are used for model training"? In general, a much more thorough description of the approach used in each architecture is necessary to understand what the authors have done.
- The results of the different models are not shown, we are directly shown the validation. How do the different models behave in the spatial and temporal distribution of MHWs? There is no mention of MHW classification (moderate, strong, severe), has this been done?
- Accuracy is not described. But even if a description is provided, it is very unlikely that a single error measure (in addition, averaged over the full domain of study) is able to provide the full complexity of model performance in forecasting MHWs at different lead times. This needs a more developed evaluation approach, also considering that the events being addressed are extreme events and need specific statistical treatment. Figures 2 and 3 are insufficient to fully assess this.
- It is not clear how the same models should be able to predict an extreme event at a few days horizon and up to 2 years of lead time. The authors have mentioned to the other reviewer that they changed ERA-5 to "ECMWF forecast", but there are several forecast systems available at ECMWF, please clarify.
- Figures 4 and 5 do not allow to assess the capacity of each model to forecast MHWs, other than their seasonal behaviour (more MHWs in summer months) and the lack of seasonality in some models as MLP. Minimising a full field of MHW distribution to a single dot is not very informative. The sentence "Most model predictions generally align with the distribution of actual MHWs" (line 334) is simply not true.
- Conclusions are not supported by the text, specially when it is stated "In both lead times, the machine learning models demonstrated superior predictive skill compared with baseline methods, achieving performance comparable to or exceeding that of other climate simulation models." This is not true.
Here are, in addition, a few minor comments:
- Title is lacking a noun defining the subject of the sentence: "An interpretable machine learning {approach, architecture...} for marine heatwave..."- Figure 1 is not very informative, other than in showing the region of study. The rest looks more as a graphical abstract, but provides no information needed to understand the paper contents.
- Acronyms should be described the first time they are used, but without considering the abstract. So SCS should be reintroduced in the introduction (even if it is introduced in the abstract) and LSTM, MLP etc should be already described in the introduction, and not only on section 2.
- Section 2.4 what are C and epsilon? introduce.
- Section 2.6: Main spatial modes: how many are these? How are these determined? Are these the same number for each variable?
- The period 1950 - 2014 is mentioned in section 3.1 for the data range, but it is unclear if this is the range used for the training as well.
- Line 254-355, not a proper sentence.
- Line 250 Describe SHAP here (it is described later)
- Line 256: what is multi-duration testing and evaluation?
- Figure 6: labels are too small to be read. Explain all labels and axes.
Citation: https://doi.org/10.5194/egusphere-2026-262-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,061 | 738 | 133 | 1,932 | 86 | 121 |
- HTML: 1,061
- PDF: 738
- XML: 133
- Total: 1,932
- BibTeX: 86
- EndNote: 121
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This study applies Machine and Deep Learning techniques to forecast marine heatwaves (MHWs) in the South China Sea. Training is performed by linking CMIP6 model MHWs to EOFs of atmospheric and oceanic predictors. Forecasts of a recent period are then made using ERA5 atmospheric reanalysis predictors and validated against satellite sea surface temperature.
The study suffers from fundamental gaps in the description and justification of the methods, and flaws in the forecast analysis. There is potential for an interesting work on using climate simulations to predict ocean extremes but major revisions and additions are first required.
Firstly, short-term forecasting and seasonal/multi-year are fundamentally different. Generally, we cannot and do not need to predict the exact state of the climate 730 days in advance. I would suggest the authors study carefully the conventions of forecasting at different timescales, and to consider the target timescale most suited to their method.
Do the predictions presented represent an area-averaged time series? The SCS is a large area with many studies showing distinct MHW dynamics in the area. I am concerned that an area-averaged time series is not truly representative of individual events, nor relevant to impacts. Such an approach should at least cover sub-domains, if not individual grid cells on a common grid.
The forecast validation (Figure 2 and the corresponding analysis) is very unclear. This is unfortunate because it is the foundation of the paper. What does it mean to have a prediction accuracy of CMIP6 models or of ML models alone? Also, ERA5 is not used as a prediction.
Accuracy is not defined.
A 7-day moving average is applied to target data. This removes variability that would be crucial for short-term forecasts, which are within the scope of this study. Please check the sensitivity of this approach.
It is not clear why Random Forest is considered the best performing model and subsequently used for the explainability analysis.
Line 202-204: Provide References.