MTGL40-5: A Multi-Temporal Dataset for Remote Sensing Image Geo-Localization

Round 1

Reviewer 1 Report

In the manuscript titled "MTGL40-5: A Multi-Temporal Dataset For Remote Sensing Image Geo-Localization", the authors collected and built a remote sensing image dataset with a temporal span. They employed relevant methods to demonstrate the necessity of this dataset for geolocation purposes. This study contains several interesting findings and highlights the shortcomings of existing geolocation methods on cross-temporal datasets. The main flaw of the article is that it fails to propose a strategy for addressing the challenge of temporal span, only demonstrating the limited applicability of many current geolocation methods on multi-temporal datasets.

1. The article, in line 106 on page 3 (fourth main contribution of the article), discusses experiments on image geolocation using advanced methods, and Table 3 presents the performance of these methods on a multi-temporal dataset. However, LCM, LPN, and RK-Ne are all advanced methods designed for cross-view image geolocalization. The question arises whether they are suitable for single-view, multi-temporal datasets.

2. The author's description of Table 1 is not specific enough, such as the size of the training set.

3. During the experimental phase, a deep learning network model was used to extract deep features from images. This method lacks interpretability and typically requires a large number of images for training. The article does not specify the exact number of images in the MTGL40-5 dataset, so it is unclear if it meets the requirements for deep learning training samples.

4. In the abstract, the author mentions that researchers often overlook the time span when collecting images to build datasets. However, existing datasets like GSV-CITIES and SF-XL do include time span information, providing rich temporal variations. What are the differences or advantages of the MTGL40-5 dataset compared to these datasets?

5. The conclusion section needs to be revised. The conclusion is the expression of the research content and author's viewpoint. The sentence 'To demonstrate...' at line 440 is not suitable for the conclusion and requires adjustment in its expression.

The English expression is fluent, with only a few sentences having issues in their expression, without any grammar problems or spelling errors.

Author Response

Thank the reviewer for his/her constructive comments. We have provided detailed responses to each point raised. For more information, please refer to the attached material.

Author Response File: Author Response.pdf

Reviewer 2 Report

First of all, I think this work is important. Actually, the dataset-related works tend to be impactful because they offer new benchmarks to test many follow-up algorthms. Besides, other follow-up researches may also derive insightful findings from the collected data sets.

 

Regarding the novelty, I feel this paper is not an outstanding case. Most of the paper content is explaining the composition and main characteristcs of the data set, and the authors also spent much effort describing the related works. The evaluation section follows the standard workflow, which runs the benchmark tests on several well-known ML models. Overall the paper writting is okay and organization is very clear. Of course the contribution of the paper is mainly due to the proposal of the data set and I will not criticize that as the normal research paper.

 

In general, I am okay to have the paper accepted. However, to make the paper contribution convincing as well as generate long-term impact, I suggest the authors open their data sets on github and attach the link in the paper.

 

 

 

Author Response

Thank the reviewer for his/her constructive comments. We have provided detailed responses to each point raised. For more information, please refer to the attached material.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper addresses an important and yet explored task of long-term geo-localization. The author have proposed an important dataset for the long-term image geo-localization task and proposed a reasonable mehtod for precise geo-localization. It is well written but some minor issues should be considered before publication.

(1) Is there a public link for releasing the proposed dataset?

(2) The author has shown some samples at different times in the dataset in Fig. 1 and 2, but somen samples at different times still have high similarity, which may cause overfitting when training the model. Can you provide more detailed examples or further explain this issue?

(3) In Fig. 4, Can you further explain why the point with highest density is selected as the coordinate for the query?

(4) Spilting the original sample into more patches can significantly improve the recall rate. If it is more than 4096, can 2 higher recall rate be obtained? Please further explain it.   

 

 

 

 

Well written, only minor careful edit of English required.

Author Response

Thank the reviewer for his/her constructive comments. We have provided detailed responses to each point raised. For more information, please refer to the attached material.

Author Response File: Author Response.pdf