Review Reports - Global–Local Feature Fusion Network for Remote Sensing Image Change Detection in Open-Pit Mining Areas

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

he paper proposes a novel remote sensing image change detection framework termed GLMECD-Net, which aims to improve change detection performance in open-pit mining areas by integrating global contextual modeling and local feature enhancement within a unified Siamese architecture. Specifically, the authors introduce a Global–Local Feature Mixing Embedding (GLME) module to jointly capture long-range dependencies and local spatial details, and further design a multi-scale cross-fusion module to enhance cross-temporal feature interaction and boundary delineation. Experimental results on mining-area datasets demonstrate the effectiveness of the proposed method for detecting complex and subtle changes.

Overall, the topic is meaningful and relevant to remote sensing change detection in mining scenarios, and the manuscript presents a potentially promising framework for ecological monitoring and mining supervision. However, although the proposed method shows certain novelty, the current version of the manuscript still suffers from several issues regarding motivation clarity, methodological rigor, experimental completeness, and writing consistency. Therefore, substantial revision is required before the manuscript can be considered for publication.

The main concerns are summarized as follows:

The Introduction lacks sufficient critical analysis and fails to establish a rigorous motivation chain. It mainly lists previous studies descriptively without in-depth discussion. Specifically, the manuscript does not clearly explain why existing methods are unsuitable for mining-area change detection, nor sufficiently analyze the limitations of different categories of methods in such complex environments. Moreover, the motivations behind the proposed architecture are described mainly in an empirical manner without adequate theoretical or mechanistic support. For example, the manuscript does not explain why “irregular boundaries” specifically require local modeling, why “fragmented regions” benefit from global contextual modeling, or why existing attention mechanisms are insufficient. In addition, the listed contributions largely repeat the abstract and lack sufficiently concrete technical descriptions. The authors are encouraged to strengthen the Introduction with deeper analysis and clearer motivation justification.
The literature review in the introduction section needs to be further refined. Moreover, the introduction is not enough, more SAR change detection methods should be introduced,such as Light-weight modality compensation network (LMCNet), Multi-scale rotation- invariant haar-like feature integrated CNN (MSRIHL-CNN), and multi-kernel-size feature fusion based convolutional neural network (MKSFF-CNN).
The methodological description in Section 3.2 contains redundancy and lacks reproducibility details. Equations (2) and (3) appear repetitive and both describe the residual MLP process within the global branch, which may confuse readers. In addition, the “Token Mixing (TM)” operator is introduced only symbolically without sufficient implementation details. The manuscript does not explain whether “TM” is based on self-attention, MLP-Mixer, convolution, or another structure, nor how it differs from existing token mixing strategies. This may affect reproducibility. The authors are encouraged to provide clearer descriptions, pseudo-code, or structural illustrations. Moreover, inconsistent equation citation styles such as “In Eq. 2” and “In Equation 3” should be unified.
The design rationale of the multi-scale cross-fusion module requires further justification. In Section 3.3, the module sequentially integrates “Multi-head Cross Attention,” asymmetric convolution, channel attention, and “Pixel Attention.” However, the manuscript does not sufficiently explain why all these mechanisms are jointly required or how the functional responsibilities of each sub-module differ. Potential functional overlap or redundancy is not discussed. The authors should clarify the motivation, necessity, and cooperative relationships of each component, and provide finer-grained ablation studies to validate their independent contributions.
There are inconsistencies between the reported experimental results in the abstract and those presented in the experimental section. The abstract states that the proposed method achieves “95.6% Precision, 89.2% Recall, 92.3% F1-score, and 85.7% IoU,” but these values do not appear anywhere in Section 4. Neither the MACD nor MineNetCD tables report matching results, leading to inconsistency between the abstract and main text. In addition, Section 4.2 states that six evaluation metrics are adopted, namely Precision, Recall, F1-score, IoU, OA, and Kappa coefficient, yet only four metrics are reported in the experiments, omitting Recall and Kappa coefficient. These inconsistencies should be corrected.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents a new type of network for detecting changes in images of pit mining images which will be highly useful for those working in the mining sectors. The structure of the network is reasonable and there is a good justification for all the structures in the network which detects the changes. For example, the use of Transformer/Encoders with Cross-Attention systems using Convolutional Neural Networks (CNN) is a good combination of the latest innovations in deep learning systems for image processing. The results of the paper show that the proposed network can provide good results compared to the prior state-of-the-art methods for such change detection.

I do have the following comments about the paper:

All dimensions and numerical parameters of the components of the network should be clearly identified in the paper. I believe that all the values are given in the paper but they are distributed all over the paper. It would be best to have all the parameters given in a single table which would help others to replicate the results.
The parameters of the training algorithm are given but it would be best to have more information on the split between training, testing, and (if used) the validation set used to find the hyper-parameters of the network. This split should be described.
It appears that the described network does generalize well so the authors should describe the method(s) used, if any, were used to regularize the network.

In conclusion, I think there is good work here. The authors just need to provide some information on the training and testing split that was used and how the methods used to ensure generalization.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript proposes Global–Local Multi-scale Cross-fusion Enhanced Change Detection Network (GLMECD-Net) to detect remote sensing image changes in open-pit mining areas. Specifically, the authors utilize a Siamese encoder and a Global-Local Feature Mixing Embedding (GLME) module to capture both long-range contextual information and local spatial details. Through experimental validation on mining-area datasets, the manuscript claims that the proposed framework provides effective and robust performance for detecting complex and subtle changes in open-pit mining areas. However, the reviewer has identified several limitations and concerns:

Although Section 2 cites several related studies regarding CNN and Transformer developments for change detection, the discussion is largely a sequential listing of what each work has done. Because the cited works are not conceptually connected, the general limitations summarized at the end feel abrupt. These limitations are unsubstantially asserted without a clear explanation of how they derive from specific architectural flaws in the cited literature. The authors should reorganize this section to synthesize the common threads and architectural bottlenecks.
At the end of Section 2, the authors summarize three key bottlenecks faced by current change detection methods. However, the manuscript fails to clearly establish how the proposed framework explicitly addresses each of these specific bottlenecks within the first 2 sections. The authors are strongly encouraged to explicitly map their proposed contributions to these identified issues.
In Figure 4, the text labels within several module blocks are disproportionately small, leaving excessive whitespace.
Although Section 5.4 acknowledges that the computational complexity of the model remains relatively high, the exact number of model parameters and FLOPs are never explicitly reported in the experimental section. This essential efficiency information should be explicitly reported to allow for a fair comparison against baseline models.
Figures 5 and 6 currently present qualitative visual comparisons. However, the subplots contain highly complex scenes, making it difficult for readers to visually verify the model's superiority. The authors are encouraged to reorganize these visual comparisons into a more informative layout.
It is better to include and comment on more state-of-the-art studies, e.g., https://doi.org/10.3390/rs16224263