Review of "Enhancing coastal winds and surface ocean currents with deep learning for short-term wave forecasting", manuscript egusphere-2025-657

This manuscript presents a practical approach to improving the performance of numerical wave models by correcting their forcing fields—namely wind and surface currents—using Artificial Neural Networks trained on remote sensing data such as SAR and HFR. The methodology is applied and validated at multiple pilot sites, demonstrating consistent and significant improvements across several key metrics. The corrected forcings lead to better wave height and period predictions, both under normal and extreme conditions. Overall, this work is methodologically sound, relevant to the field of operational ocean forecasting, and contributes meaningful advancements in the integration of remote sensing with data-driven modeling techniques. Therefore, after making some appropriate revisions (mainly formatting issues), I believe this manuscript is suitable for publication. Here are some of my comments about the manuscript.

Major comments:

1. I noticed that a GAN-based architecture was used for wind field correction, while an autoencoder-like structure was adopted for surface current correction. Could the authors comment on the rationale behind selecting these different architectures for the two tasks? Also, were other model types explored or compared during the development process, e.g., if a CNN-based super-resolution network was used directly instead of a GAN model (i.e., SAR is used directly as a target, with the network output calculating an RMSE-like loss), would this be any less effective? (Note that there is no need for the authors to add additional experiments here, just a brief discussion)
2. Figures 2 and 3 clearly illustrate the model architectures, and they are generally well-presented. However, the diagrams could be further improved by including more detailed information on the data dimensionality. For instance, adding the input and output shapes at the beginning and end of each model—either directly in the figures or in the accompanying text—would help readers better understand how the data is transformed through the network. This additional context would make the architecture more transparent and informative, especially for those interested in replicating or adapting the models.
3. I suggest the authors include a brief subsection (such as 2.6 Error Metrics) in Section 2 that summarizes all the error metrics used throughout the manuscript. This summary should provide the definitions and explicit formulas for each metric (e.g., RMSD, bias, correlation, etc.). Doing so would enhance clarity and help readers better understand the evaluation criteria, especially those who may not be familiar with all the statistical indicators applied.

Detailed comments:

1. The authors should pay close attention to citation formatting throughout the manuscript. For instance, in Line 70, the citation "(Gurgel et al., 1999)" is correctly formatted, but in other places (e.g., Line 52: "Hauser et al. 2023"), the comma after "et al." is missing. Such inconsistencies should be carefully checked and corrected. Additionally, figure references should follow the format of the journal—"Figure X" is appropriate at the beginning of a sentence, while "Fig. X" should be used elsewhere. Some citations are also inconsistently bolded, which should be standardized to maintain uniform formatting. Issues like this hopefully the authors can address them in a revised manuscript
2. In Line 188, the authors refer to "training/validation datasets" in the context of evaluating model performance. However, if the dataset mentioned here is used solely for post-training evaluation rather than during model training for purposes like early stopping or hyperparameter tuning, it would be more accurate to refer to it as a "test dataset" rather than a "validation dataset". Similarly, the term "validation period" used later in the manuscript should be revised to "test period" or "evaluation period" which may help avoid confusion.
3. In Lines 301–314, multiple date formats are used inconsistently, such as "January 2021 – January 2023", "Sep 2021 – Jan 2022", "25^th – 27^th November", "November 26–27, 2021", and "20th–23rd January 2022". I recommend standardizing the date format throughout the manuscript for consistency and improved readability.
4. Many formatting inconsistencies can be noted in Fig. 5 and Fig.9. For example, for the scatterplot, while the scale intervals are numerically the same, the x-axis has a sparser scale density than the y-axis. Also, the gridlines are either present or absent. The unit notation is also different between the two: one uses ‘[m/s]’ while the other uses ‘(m/s)’. In addition, the 1:1 reference line is drawn in different colors - red on the left and green on the right - which may cause unnecessary distractions. Standardizing these visual elements will enhance the overall coherence and presentation quality of the charts.
5. In Fig. 7, which displays both positive and negative deviations, I suggest adjusting the color bar so that the central (white) point is explicitly labeled as 0. Additionally, using symmetric tick values for positive and negative ranges—ideally with a limited number of decimal places (e.g., [..., –7.6, –3.8, 0.0, 3.8, 7.6, ...])—would improve both the readability and the aesthetic quality of the figure. Also, the word spacing in the subheading of this image is odd.
6. In 13–16, there is a noticeable mismatch in the color tone between the plot lines and their corresponding legend entries—for example, while both may be shades of blue, one appears significantly lighter or darker than the other. If the legends were added during figure post-processing, using a color picker tool to precisely match the tones would improve the visual coherence. Although this does not affect the scientific interpretation, ensuring consistency in color tones would enhance the professionalism and clarity of the figures.
7. Throughout the manuscript, there are noticeable inconsistencies in figure formatting that should be addressed. For example, multiple styles are used for subfigure labels, including (1), [1], and (i), which creates confusion and detracts from the overall professionalism. Additionally, figure and table titles vary in formatting—some are in italics while others are in regular font, which should be standardized. Moreover, the resolution of several figures appears to be quite low, with visibly pixelated text and labels.

This paper describes a method for adjusting operational surface fields, specifically wind speed and current speed, to minimise forecast errors in the coastal zone. Artificial neural networks trained with radar data from land stations (HFR) and satellites (SAR) adjust forcing fields. Overall, the paper reads well and contains all relevant references. However, the vast amount of acronyms limits readability to some extent. The methodology is reasonably well documented, and the verification of surface winds and currents appears sound. I appreciate the attempt to separate errors in the forcings that drive the wave model.

Verification for "short-term wave forecasting" is missing. Perhaps this is out of scope, then the title would need to be edited accordingly. Wave verification is too brief and could benefit from a more detailed analysis, other than a case study. Time series plots presented in Figure 13-16 did not convince me of an improved forecast system. Numerical weather prediction models feature some variability in skill over a complete seasonal cycle. Training and verification for a particular season can introduce seasonal bias in the correction method.

The introduction could benefit from a paragraph on the expected impact on the wave model skill, not just for the wind generation part, but also for the effect of currents on wave predictability.

Specific comments

Referencing in text uses inconsistent punctuation and contains et al. year, et al., year and author, year.

Line 29-30: "Third generation spectral wave models ..., as they address wave generation and propagation". Third-generation models explicitly represent the nonlinear wave-wave interactions. Generation and propagation refer to second-generation models.

Line 31-32: "three factors". What about the representation error? In the coastal zone, many processes are often overlooked, such as wetting and drying, river discharge, variable water depth, changes in beach profile, and seagrass, among others.

Line 34: "(ii) reduce biases and errors in the inputs (Durrant et al 2013) …". Complementary to Durrant et al., Zieger (2025) assessed forecast errors in an operational wave model. (https://doi.org/10.1071/ES25010).

Line 38-39: "Reducing biases and errors in the forcings have certain advantages...it does not require changing the physics." The parameterisations that represent the physics will naturally contain some form of bias that stems from model development. The underlying tuning parameters are bound to the frequency of model fields (daily, 6 hours, 3 hours, hourly) and the underlying statistic (i.e., cumulative, average, instantaneous). As a result, one has to change the physics kind of.

Line 84: "spectacular growth" does not sound well. Perhaps "exponential growth".

Line 90: "concrete". This word does not feel right and does not add clarity to the sentence (omit). There is a slight overuse of the word 'concrete' in the manuscript.

Line 128: Could you please provide more specific details on the ST4 configuration? Could you state the value for BETAMAX used?

Line 186-187: "It has been discarded images" is gibberish. Please rephrase the sentence.

Line 360-361: "So, it can be concluded that the ANN is able to predict winds that are closer to the SAR data than IFS." Would one not verify against an independent dataset (also Figure 5)? The ANN training would be well aware of the characteristics and structure of SAR data.

Line 479: "Benefits from the ANN forcing are more remarkable during extreme events. During storm Arwen, strong Northern winds at the Northern Iberian Peninsula (Figure 12a) …" What is the rationale for showing daily-average marine wind speed to depict extreme events? The most pronounced feature in wind speed difference (Fig. 12b) is the north-west south-east shift around coastlines. In addition, wind speed verification at the GAL station (Fig. 13 and Fig. 14) indicates that surface wind speed is consistently overpredicted. At the same time, the responding wave field exhibits lower magnitudes for the duration of the storms. On the Mediterranean side, GCA station, is not able to capture the double peak on 19^th and 20^th January. Can you elaborate on this?

Line 671: "The ANN forcings have positive impact on the wave forecast, especially under extreme events." To be honest, the results for significant wave height seem to reduce the variability in the signal. Peaks are less pronounced in Figure 15, and almost no signal in the time series in Figure 16. In general, a diurnal signal appears to be present in the observed wave field, which is lost in the simulation.

Line 674: "Wind ANN tends to decrease the overestimation of ECMWF-IFS wind speed". Is this correct? Time series plots in Figure 13-16 indicate that ANN winds are frequently higher than ECMWF and observations.