A Complete YOLO-Based Ship Detection Method for Thermal Infrared Remote Sensing Images under Complex Backgrounds
Round 1
Reviewer 1 Report
The manuscript has been improved as requested.
Author Response
Thank you very much for all the reviewers' opinions. Your advice means a lot to us.
Reviewer 2 Report
Reviewer’s Report on the manuscript entitled:
A Complete YOLO-based Ship Detection Method for Thermal Infrared Remote Sensing Images under Complex Backgrounds
The authors proposed a complete YOLO-based ship detection method for TIRSIs under complex backgrounds. The authors have addressed my earlier comments. The length of the manuscript is short for an “Article” type. At the present form it should be “Technical Note” type.
Please also find below some additional comments.
Line 207. This section should be called: Materials and Methods
Lines 220, 221. Please replace “nearest interpolation, bilinear interpolation, and bicubic interpolation. [24]” with “nearest, bilinear and bicubic interpolations, and spectral techniques [24].”
The Discussion section is short. Please add a few paragraphs and discuss the results in the light of other similar studies and also mention the limitation of the research.
Finally, please carefully proofread the manuscript
Thank you for your contribution
Regards,
Author Response
Author response to Reviewer#2: Thank you for the serious reviews. The type of the manuscript is changed to “Technical Note”. The section is changed to “Materials and Methods” in Line 207. In lines 220, 221, “nearest interpolation, bilinear interpolation, and bicubic interpolation. [24]” have been replaced by “nearest, bilinear and bicubic interpolations, and spectral techniques [24].” The limitation of the research and discussion of other similar studies are added in Discussion section, as following “5. Discussion: Visible remote sensing images have the advantage of high resolution and have been widely used in ship detection. However, visible RSIs are based on the reflection of light, and it is difficult to see and identify targets in the completely dark environment or the condition that the light is not enough. SAR technology can work all day and has a long detection range. However, radar observation is susceptible to interference from echoes of waves, islands and land, radio frequency and atmospheric noise, which makes it difficult to detect targets on the sea surface. By measuring changes in infrared radiation caused by differences in targets temperature and radiation, infrared thermal imaging converts invisible infrared light into visible content. It has special application value in hotspot area monitoring, camouflage target disclosure, and military target detection, so it is paid attention to by the major military powers. Compared with visible light, infrared has the advantages of strong smoke penetration and all-day work. Different from SAR, infrared imaging passively receives radiation, with good concealment and stronger security. Therefore, targets detection of infrared images have excellent applicability in complex sea conditions.
Our approach is based on supervised training. High quality, balanced, standardized, and thoroughly cleaned datasets are required, otherwise, it will lead to poor results of supervised learning. In the 30m resolution TI images, the geometric features of the ship close to the river bank are very different from those of the independent ship in the middle of the river or sea. The aspect ratio of different types of ships are summarized and calculated by literature research. According to the above research, in the process of making the datasets, only ships with an aspect ratio of 4.23-7.53 are annotated. This was not noticed in our initial work, and based on the datasets originally annotated, the IC-CYSDM model are obtained. To increase the reliability of datasets, the aspect ratio of the labeled ships are carefully chosen to obtain the final datasets, which are used to train CYSDM model. Through the experimental comparison, the large difference of ship targets aspect ratio, namely the large intra-class variation, lead to the decrease of the networks detection efficiency. The experimental results after eliminating intra-class variation show that the proposed method (CYSDM) is suitable for ship detection in complex scenes including sea, bay, river, and covered by cloud. And it provides a reference for sensitive TI ship targets search in all-day. However, the detection results of the proposed method are poor for ships close to the shore and adjacent vessels. To solve the above problems, multi-frame information will be utilized in the future work.”
Thank you very much for all the reviewers' opinions. Your advice means a lot to us.
Author Response File: 
Author Response.pdf
Reviewer 3 Report
My comments have been addressed
Author Response
Thank you very much for all the reviewers' opinions. Your advice means a lot to us.
Reviewer 4 Report
The authors have been complied with the suggestions made.
I suggest to publish the manuscript.
Author Response
Thank you very much for all the reviewers' opinions. Your advice means a lot to us.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
Why is that specific intraclass variation chosen. Generally the ship length/width ratio is just around 7. It is counter intuitive later, that the IC-CYSDM algorithm is the one without intraclass.
This work describe a CYSDM YOLO algorithm for ship detection in therma IR remote sensing river images from SDGSAT-1. It shows promising results which, however, need better clarifications.
The algorithm is poorly described. Fig.1 is rather complicated and not explained. It is not clear how the YOLO algorithm works. Is it a sliding window or what?
In Table 1 it is not clear which datasets that the bands and resolutions belong to.
In line 190 it is mentioned that google map ships are used for annotation. How is that done, and how many ships are annotated? This DB is the most important element for training. The precise number of ships should be given in table 1 for TP, TN,FP,FN.
The notation should be cleaned. Fx Precision and Recall are several places shortened to Pr and Re, and some places P-R.
Eqs. 2-6 introduce a number of variables that are not defined.
It should be explained in table 2, why the YOLO algorithms vary in speed by up to a factor 70.
I could not find a definition of mAP for fig. 3.
It would also be nice with a description of how the IR images improve classification as opposed to optical or SAR image classificaiton.
In summary, this work cannot be recommended for publication in its present form. Basic description of the method, algorithm, annotated database, etc. are missing.
Reviewer 2 Report
Reviewer’s Report on the manuscript entitled:
A Complete YOLO-based Ship Detection Method for Thermal Infrared Remote Sensing Images under Complex Backgrounds
The authors proposed a YOLO-based ship detection method for TIRSIs under complex backgrounds. I found the topic and results interesting. The presentation can be further improved. Please find below my comments.
Line 18. Grammar Issue: Please remove “a”
Lines 19, 23. Please define ALL the acronyms. Please ensure that all the acronyms are defined the first time they are defined and be consistent with their style. Please also add an acronym table at the end of the manuscript listing all the acronyms used in the manuscript.
Line 20-21. Up-sampling from 30m o 10m is a limitation. Questions may rise on why bicubic interpolation why not other techniques what happens with aliasing, etc. Therefore, I suggest the authors remove this line from the Abstract. This, however, should be mentioned in the Methodology and in the Discussion part as a limitation.
In Section 2 (Previous Related Research), please add another paragraph and include the following most recent references as well:
Ship detection using self-supervised learning:
https://doi.org/10.3390/rs13214255
Ship detection in SAR images based on multi-scale feature extraction and adaptive feature fusion:
https://doi.org/10.3390/rs14030755
A review on change/target detection techniques using time series, spectral, and wavelet analysis:
https://doi.org/10.3390/app11136141
Please add another subsection at the beginning of Section 3 for the Study region. Please show a geographic map with latitude, longitude, and scale bar, and briefly talk about the typical atmospheric condition for that region that may affect the imagery.
Lines 249, 251, 254, etc. It is preferred to use the word “Equation” instead of “Formula”. Please check and replace “Formula” with “Equation” everywhere.
I suggest authors add a Discussion section before the conclusion to discuss the results in the light of other similar research and also the challenges and recommendations.
Please check the references one by one to ensure their correctness. Also, please follow the MDPI guideline for the style and format of the references.
Thank you for your contribution
Regards,
Reviewer 3 Report
The authors have proposed a YOLO based framework for detecting ships from Thermal Images.
1) The presented experimental results are not at all sufficient. The authors need to compare their approach with other Deep Learning detection methods [1] and other traditional approaches as described in [1 - 3]
2) References need to significantly improve. Please look at prior detection works done in ISAR and SAR images which are also similar to Thermal Images
3) The paper currently lacks novelty and needs to significantly improve
[1] Chen, C., Wang, B., Lu, C.X., Trigoni, N. and Markham, A., 2020. A survey on deep learning for localization and mapping: Towards the age of spatial machine intelligence. arXiv preprint arXiv:2006.12567.
[2] Theagarajan, R., Bhanu, B., Erpek, T., Hue, Y.K., Schwieterman, R., Davaslioglu, K., Shi, Y. and Sagduyu, Y.E., 2020. Integrating deep learning-based data driven and model-based approaches for inverse synthetic aperture radar target recognition. Optical Engineering, 59(5), p.051407.
[3] B. Xue and N. Tong, “Real-world ISAR object recognition using deep multimodal relation learning,” IEEE Trans. Cybern. (2019).
[4] B. Xue and N. Tong, “DIOD: fast and efficient weakly semi-supervised deep complex ISAR object detection,” IEEE Trans. Cybern. (2018).
Reviewer 4 Report
It is a well written manuscript in the field of "thermal infrared remote sensing".
The Introduction and the previous related work sections are well substantiated.
The necessary references are updated.
The proposed method and the used databases are extensively explained.
My opinion is that it should be published with minor revisions:
1. Initials should be explained in the first place they appear.
2. The conclusions are quite brief. More material should be added regarding the comparisons between existing techniques and the proposed method.
