Dynamic Loss Reweighting Method Based on Cumulative Classification Scores for Long-Tailed Remote Sensing Image Classification

Round 1

Reviewer 1 Report

This paper proposes a novel loss reweighting method to addresses the long-tailed problem of remote sensing image classification. In this paper, the authors widely analyze some shortcomings of the current remote sensing image classification and point out the fact that long-tailed remote sensing datasets are very common. Unlike the traditional loss-weighting method based on the number of samples from each category, the authors propose a dynamic loss-weighting method based on the category cumulative classification scores, which is a good solution to this problem. In sum, the proposed method is feasible and interesting. Also, the manuscript is written and organized well. Only some minor suggestions are as follows.

1.     The figures in the article are not clear enough, and the quality of the figures needs to be improved.

2.     Are there some errors in Equation 13? Should R^pos_2c be replaced with R^neg_2c?

3.     The paper mentions that static indicators cannot be adapted to the dynamic training process of the model, but it does not elaborate on what results this situation leads to. Why is it not as good to use static indicators as dynamic ones?

4.     There are two ways to calculate classification networks, the “softmax” and the “sigmoid” functions. The paper should explain under what circumstances these two functions are used to calculate the classification scores, and how this corresponds to the balanced cross-entropy function that follows?

5.     There are many methods to solve the long tail problem, such as “SSD”, “OLTR”. what is the advantage of the method proposed in the paper compared to these methods? Are there flaws and shortcomings in these methods? Refer to:

[1] Z. Liu, Z. Miao, X. Zhan, J. Wang, B. Gong, and S. X. Yu, “Largescale long-tailed recognition in an open world,” in Computer Vision and Pattern Recognition, 2019, pp. 2537–2546.

[2] T. Li, L. Wang, and G. Wu, “Self supervision to distillation for long-tailed visual recognition,” in International Conference on Computer Vision, 2021.

[3] X. Chen, J. Jiang, Z. Li, H. Qi, Q. Li, J. Liu, L. Zheng, M. Liu, and Y. Deng, “An online continual object detector on VHR remote sensing images with class imbalance,” Engineering Applications of Artificial Intelligence, vol. 117, pp. 105549, 2023/01/01/, 2023.

[4] X. Chen, Z. Li, J. Jiang, Z. Han, S. Deng, Z. Li, T. Fang, H. Huo, Q. Li, M. Liu, “Adaptive Effective Receptive Field Convolution for Semantic Segmentation of VHR Remote Sensing Images, “ IEEE Transactions on Geoscience and Remote Sensing, 59 (2021) 3532-3546

6.     The loss reweighting method based on cumulative classification scores is prone to unstable training. How does the proposed method in the paper ensure that the network can be trained smoothly?

Author Response

We thanks a lot for reviewer. The responses item by item are in the attachment, please check it. 

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper addresses the remote sensing  image classification problem. The approach is built upon CNN feature leaning framework. The motivation is to solve the long-tailed problem which widely existed in the related dataset. The paper is overall well written. Techniques introduced is solid and reasonable to solve the long-tailed problem. Experimental results show the effective of the proposed method. There are some minor problems in the manuscript. The proposed sample reweighing strategy is somewhat similar to the idea of focal loss introduced in the domain of object detection. Suggest to cite the paper and make some comparisons in both idea and implementation aspects.  In the experimental part, there is no introduction about other comparison methods. Also it's better to make some deep analysis about why the proposed method can achieve better performance than the other SOTA methods.

Author Response

We thanks a lot for reviewer. The responses item by item are in the attachment, please check it. 

Author Response File: Author Response.pdf

Reviewer 3 Report

This the review report of a paper titled “

Dynamic Loss Weighting Method Based on Cumulative Classification Scores for Long-tailed Remote Sensing image Classification”

General comments

1. The paper proposes a novel method, “category weight Matrix based on cumulative classification score,” that is described elaborately in section 3.2

2. The paper proposes an Equilibrium Cross-entropy Loss Function in section 3.3 for the loss reweighting method based on category cumulative classification scores.

1. The proposed method (1 and 2) above is evaluated with four datasets (SIRI-WHU, NWPU-RESISC45, PatternNet, and AID) and experiment on the long-tailed datasets with imbalance ratios of 0.01, 0.02, and 0.05 demonstrate the superiority and robustness

3. of proposed method compared to the state-of-the-art techniques in the literature.

4. The paper is structured coherently.

Specific comments

1. In section 5.2, the parmeter analysis typo needs to be corrected.

2. I strongly recommend this paper for publication.

Comments for author File: Comments.pdf

Author Response

We thanks a lot for reviewer. The responses item by item are in the attachment, please check it. 

Author Response File: Author Response.pdf

Reviewer 4 Report

Paper: Dynamic Loss Weighting Method Based on Cumulative Classification Scores for Long-tailed Remote Sensing image Classification

  Comments: 1. Authors are claiming that the proposed method for remote sensing image classification in the title, but they have tested only RGB images. 2. Information about the datasets are not uniform. In some cases, spatial resolutions are given and in some cases are not given.  3. Methodology section should be rewritten, and repetition should be minimised. 4. Training and optimization of the models are not discussed. Did you use the same hyperparameters for all the models? What is the basis of hyperparameter tuning?
5.  I have serious doubts if you are claiming that the proposed method performed well on a few sample datasets. But I can see results are good on  PatternNet dataset. 6. Why did you use an imbalance factor for SIRI-WHU, NWPU-RESISC45 and PatternNet datasets and a decay factor for the AID dataset? 7. Why did you not compare your results on the latest long-tailed problem related methods? eg.  a)  'Duan, Y.; Liu, X.; Jatowt, A.; Yu, H.-t.; Lynden, S.; Kim, K.-S.; Matono, A. Long-Tailed Graph Representation Learning via Dual Cost-Sensitive Graph Convolutional Network. Remote Sens. 2022, 14, 3295. https://doi.org/10.3390/rs14143295" b) W. Zhao, J. Liu, Y. Liu, F. Zhao, Y. He and H. Lu, "Teaching Teachers First and Then Student: Hierarchical Distillation to Improve Long-Tailed Object Recognition in Aerial Images," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-12, 2022, Art no. 5624412, doi: 10.1109/TGRS.2022.3177853. 8. There are macro mistakes in the paper, e.g.In the conclusion, "Based on these, Based on this,". Author should be more specific while writing the paper.    

Author Response

We thanks a lot for reviewer. The responses item by item are in the attachment, please check it. 

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

This paper can be accepted for publication.