Ocean Satellite Data Fusion for High-Resolution Surface Current Maps

Round 1

Reviewer 1 Report

In this study, the authors develop a real-time operational pipeline for high-resolution ocean surface current reconstruction. Multimodal inputs are used to train a convolutional neural network to predict the velocity field. It is trained by the Observing System Simulation Experiment (OSSE). The authors' model exhibits consistent performance for both L3 (cloudy) and L4 (interpolated) SST data, which is a significant improvement over previous methods. In addition, our novel pseudo-labelling strategy allows us to fine-tune our model on real data, narrowing the domain gap and improving robustness for real-world applications. To validate the results in real satellite observations, sparse in-situ measurements obtained from drifters are used. The results show that the method in this paper improves the reconstruction accuracy of the current strength and direction compared to the standard method.

Overall, the manuscript is detailed, innovative in its research tools, and worthy of revision for publication. However, there are some questions and suggestions for the authors to improve and add.

1. Can the framework of Figure 5 be explained in more detail. The reviewer was interested in this figure, but did not fully understand the intermediate process.

2. The neural network model CNN is prone to overfitting in the training process. The authors do not seem to have elaborated on the adverse effects in this regard. Please add, suggesting that there are diagrams to support this.

3. CNN models need to be calibrated and tested after training, please add and elaborate on this process. Without this introduction, it is difficult to convince the reviewers of the authors' conclusions.

4.It is better to add a new section of conclusion to the article, describing the authors' contributions and conclusions in separate articles. As well as elaborate on the shortcomings and outlook of the study.

5.Insufficient introduction on machine learning in the introduction.

6.Can Figure 12 be elaborated in more detail. Because this figure is very important

Author Response

We thank the reviewer for their positive feedback on our work on predicting ocean currents, as well their helpful comments and suggestions for improvement. 

1. As requested, we add additional detail in both the caption of Figure 5 and the corresponding section of the text that explains the U-Net encoder-decoder architecture that we use as our model backbone. In particular, we detail the convolution operation used, which uses a 3x3 kernel with a ReLu activation function and extracts features from input data. Downsampling reduces spatial dimensions to capture larger context, while upsampling increases them to recover finer details. Skip connections connect corresponding layers between the encoder and decoder, aiding in the retrieval of high-resolution features and mitigating vanishing gradients. We refer to Ronneberger et al. 2015 (original U-Net paper) for further details on the U-Net architecture.
   
   
2. In order to avoid overfitting, we choose to train our model on 100,000 training examples from the CROCO numerical simulation. This number is chosen by evaluating our model on a small subset of drifters, which are not included in our test set. We find that when we train our model on additional samples, it tends to overfit to the simulated data and generalises less well to the real data. We add this information to Section 2.9.
   
   
3. Indeed, our CNN model needs to be calibrated and tested through numerous experiments. Normalisation is a key tool that we use to close the domain gap between simulated and real data. We have thus further emphasised the importance of normalisation in Section 2.9 to alleviate systematic biases which may be present between the simulated data used during training and the real satellite observations which are used during evaluation.
   
   We conduct numerous experiments on different types of inputs to our model (including L3 SST data, L4 SST data, chlorophyll) and conduct ablation studies on model architecture (number of decoders, using SSH as an output). We evaluate these experiments using our drifter test set.
   
   
4. We thank the reviewer for the suggestion to add a conclusion proceeding the discussion, and we have added this.
   
   
5. We have added further details on Figure 12 on how we compute the optimal ship route using real-time predictions of ocean surface currents, both in the caption and in the corresponding text in the Discussion section. A roll-on/roll-off cargo ship traveling along the Algerian coast in November 2023 benefited from our predictions of our ocean currents. Initially, its path was hindered by several coastal eddies, which can be accurately located using our method. We use isochrone methods to compute waypoints to propose an optimal route, which we propose one day in advance using available near real-time satellite data. With our short-term optimization the vessel altered its course slightly northward, circumventing the eddies. This minor detour allowed the ship to save over an hour on this journey and cut fuel consumption. The impact of the surface currents can be computed by subtracting measurements of the Speed Through Water (STW) from onboard instruments from the Speed Over Ground (SOG) measured by the ship's automatic identification system (AIS). Figure 12 then compares the currents felt by the ship and our ocean current predictions. 

We are attaching a version of our paper with the associated modifications. 

Author Response File: Author Response.pdf

Reviewer 2 Report

From my end, because I did similar research recently, so I could understand the goal and results in this article.
So, if you ask my opinion, I would accept this article in current form, BUT i feel that I am not qualify to edit the English for this article.

I have no  further quiz or comments on this article.

Author Response

We sincerely thank the reviewer for recognising the contributions of our paper towards real-time operational predictions of ocean currents and for their time in reviewing our article. 

We are attaching a version of our paper with minor corrections and clarifications suggested by other reviewers.

Author Response File: Author Response.pdf

Reviewer 3 Report

This is an interesting and well written paper.  It should be published.

The authors used a convolutional neural network similar to those used in computer vision to combine sea surface height and sea surface temperature from satellites to map surface currents.  The model is trained using output from a high-resolution numerical model.  Such models may not accurately represent the real ocean at a particular time, but they produce numerical oceans with the same physics as the real ocean.

The computations are evaluated by comparing model current speeds and directions to those calculated from drifting buoys.  This is a good objective method that relates to practical use.  From the stated treatment of the buoy data, it appears that the model is tuned to produce daily average currents.  If that is correct, it should be stated clearly in the text.  Most of the numerical evaluations were for the percentage of “correct” current speeds predicted within 0.15 m/sec.  Why was that metric chosen instead of using statistics of the average error or average percentage error?  The definition of “correct” current speed is buried in the caption for Table 1.  That definition should be stated more clearly in the text. 

The match of calculated current maps versus drifter paths in Figure 9 is impressive.  Is the buoy track for one day of travel?  A scale for the current vectors would be helpful, particularly if the odd units of degrees per day were used.  The qualitative maps in Figure 8 are also impressive, but they leave open the questions of how well the coherent vortices from HIRES match ground truth.  Perhaps it would be useful to show a comparison based on the validation half of the numerical model data.  That is not the same as using ground truth but it is consistent with the assumption that the numerical model creates a possible ocean.

In the caption for Figure 2, the definitions of high-resolutions and low-resolution are mixed up.

Is Figure 3 for a particular day?  What day?  The text on the scale bar is too small to read, at least on the printed copy I was using. 

Figure 9 needs labels for AVISO and HIRES at the top.   

Author Response

We thank the reviewer for their favorable review of our research concerning the prediction of ocean currents, and we appreciate their constructive recommendations for improving our work.

Indeed, the time frequency of our model predictions is daily, and we have made it clearer in multiple locations in the abstract, introduction, method and conclusion that we predict daily average currents. 

We have added further details in Section 2.7 (Evaluation Metrics) on our evaluation strategy to clarify our choice of metrics. We consider two metrics: the percentage of correctly predicted angles and percentage of correctly predicted magnitude values for our set of drifter-day observations. These simple and interpretable metrics capture the frequency of correct predictions of the direction and magnitude of the force of the currents applied to a ship or another floating object in the ocean. The threshold for correct angles is +-45 degrees and the threshold for correct magnitudes is +- 0.15m/s. As low-magnitude currents do not have a strong impact on ships, we consider only drifter observations which measure currents of over 0.25m/s during evaluation, and the average magnitude of currents measured by drifters in our test set is 0.3m/s. The threshold of 0.15m/s thus corresponds to around a 50% error in current magnitudes.

We have added further information to the caption that the drifter observations in Figure 9 are for two consecutive days, and not for one day. This is to improve visualisation of the drifter trajectory. Thank-you for picking up on this. Moreover, we have provided further information in the caption of Figure 9 of the average magnitude of the currents measured by the drifter, as a reference for the scale of the arrow depicting the drifter direction. 

In order to assess performance of our model on more complete ground truth data - albeit from numerical simulations - we visualise the magnitudes and the vorticity predicted by our model compared to the (simulated) ground truth in Figure 10. This is conducted on the validation subset of the CROCO simulation. When compared to AVISO/DUACS, our model is far superior in capturing fine-scale structures. Moreover, we provide results of our model on the simulated CROCO validation set in Table 10. 

We have made the necessary corrections to:

• Fix high-resolution vs. low-resolution mix-up in the caption of Figure 2
• Increase the size of the scale bar for Figure 3
• Add information on the dates used for the drifter visualisations in Figure 3 (2 years from 2021-2022)
• Adding labels for AVISO/DUACS and HIRES-CURRENTS in Figure 9

We are attaching a version of our paper with the associated modifications. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors responded well to the reviewers' comments and suggestions, and the revised manuscript is recommended for publication