Detection and Identification of Mesoscale Eddies in the South China Sea Based on an Artificial Neural Network Model—YOLOF and Remotely Sensed Data

Round 1

Reviewer 1 Report

Albeit the application of machine learning computer vision techniques to physical oceanography is interesting and appealing, I found the results not really too significant in oceanography and not very original in the machine learning context. It is more an extended and technical User Case for CMEMS than a scientific communication. 

Author Response

Thank you for your comments. This study aims to extract mesoscale eddies intelligently based on deep learning technique. The method of extracting mesoscale eddy features in the physical ocean to adjust the relevant parameters was adopted to improve the YOLOF target detection network in mesoscale eddy recognition research. It is an innovation of this study. Using the sea surface height anomaly data of nearly 30 years, the eddy is identified according to the corresponding rules of SLA closed contour method. The regression fitting formula between the longitude and latitude information of the eddy boundary and the coordinates of the label frame is established to automatically generate high-precision label frames. The difference between this method and the traditional method is that the ocean mesoscale eddy annotation for the latter is achieved generally through visual annotation by experts who are advanced in the field, which can induce subjective error in the detecting results. Additionally, the mesoscale eddies in the South China Sea from 1993 to 2021 were detected and identified by the proposed model. Compared with the traditional recognition methods, the model has a better recognition effect, with an accuracy of 91%. It also avoids the influence of setting threshold on the identification of mesoscale eddies and improves the detection and recognition speed to a certain extent. Therefore, not only the method but also the research conclusion in this study are meaningful for intelligent detection of mesoscale eddy.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes a YOLOF target detection model on image data. Large amount data are collected to pre-train, train, validate and test the model. A deep transfer learning method is also adopted. Results are evaluated on mAP, and compared with other traditional eddy detection model. Additionally, mesoscale eddy parameters and temporal analysis are given. More details are described in the attachment.

Summary

This paper proposes a YOLOF target detection model on image data. Large amount data are collected to pre-train, train, validate and test the model. A deep transfer learning method is also adopted. Results are evaluated on mAP, and compared with other traditional eddy detection model. Additionally, mesoscale eddy parameters and temporal analysis are given.

 

General

This is an interesting paper on an interesting topic but it lacks severely on several aspects. Some important matters are discussed only very superficially, such as the differences between YOLOF model and other models. Some claims are made that are not corroborated by the results. The localization accuracy is completely lacking from this work as well. These issues must be addressed and require major revisions.

 

Methods

1.       Figure 2 seems to be the steps of data processing, training and testing, rather than the network framework diagram in our traditional sense. A more detailed network architecture model diagram is needed.

2.       Some of the tables in the paper seem to be in different formats. For example, the border thickness of the Table 1, Table 2 and Table 3 differ, which influences the coordination。Table 4 is not the three-line table needed.

3.       Is the YOLOF model improved? Or just apply the original YOLOF model. Please illustrate clearly the improvement of the target detection model. In this paper, there is detailed description of the improved model: a Novel YOLOv3 Model With Structure and Hyperparameter Optimization for Detection of Pavement Concealed Cracks in GPR Images.

 

Results

(1)    Where are the results of the traditional eddy detection model from? Better to cite them.

(2)    Figure 4 only illustrates the results of the pre-processed samples, but what is the feature of the original samples, please display them to contrast. In this paper, the pre-processing and data augmentation methods are described: GPR-based detection of internal cracks in asphalt pavement: A combination method of DeepAugment data and object detection.

(3)    There is a mistake in Line 508 about the name of “YOLOF model”.

(4)    Is the only listed mAP index able to describe the accuracy and performance of the model? Evaluation indexes such as IoU, P and R, which are defined in the methods part, can be listed too.

(5)    Models are evaluated on recall and precision. However, with object detection the localization accuracy is also important but is not included in the analysis. A model could have high recall and high precision but general poor localisation making the model unusable for real world applications. In this remotesensing paper, above indexes are listed: Automatic Detection of Pothole Distress in Asphalt Pavement Using Improved Convolutional Neural Networks.

 

 

 

Conclusions

1)      The conclusion part looks more like a brief summary of the methodology, lack of a detailed description of the conclusion. Additionally, how is the innovation of this paper? The author combines the existing target detection model and transfer learning method for extracting mesoscale eddy features. The target detection network needs to be improved based on the existing models and combined with the characteristics of the detected objects.

 

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Please read following warning about the manuscript

Introduction

Row 33 – what do the authors mean with “fast speed” and “strong velocity”? It seems a repetition

Row 35-36 – it is not clear what means “large span”; what do the authors mean “maximum vertical depth of the deep seabed”.

Row 55 – misprint

Row 74 – a consideration about physical-parameter- based methods is given after the authors have moved on consideration about field geometry-based methods. The passage from a concept to another one, and next go back to previous concept confuses the reader

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Thank you for your detailed replies to my comments. Interesting topic and well-written paper in general. Thus, I do recommended the current version of the manuscript for publication in Remote Sensing.

Author Response

Thank you very much for your comments. Your comments are very important to us. Thanks again.