MAEANet: Multiscale Attention and Edge-Aware Siamese Network for Building Change Detection in High-Resolution Remote Sensing Images

Round 1

Reviewer 1 Report

This article presents an unprecedented building change detection framework, consisting of three parts, taking optical images only into account. It has a better performance in all evaluation parameters than the four state-of-the-art change detection methods: MSPSNet, SNUNet, STANet and EGRCNN.

The article is well written and easy to understand. It also presents all necessary references.

It would be interesting:

• if the authors could estimate the improvement of the five parameters of Table 3, if the DSM would be available, in the discussion and comment it in the conclusion;
• if the MAEANet algorithm could also improve the quality of the DSM generation in the conclusion.

Author Response

Dear Reviewer, 

Thanks for the useful comments and suggestions, please find the item-by-item response in the attached file. 

best regards, 
Xin Su

Author Response File: Author Response.pdf

Reviewer 2 Report

This study introduces a novel deep learning-based building change detection (CD) network for large areas that overcomes current limitations such as a lack of robustness to false-positive changes and confusion of boundaries of dense buildings. This proposed model called Multi-scale Attention and Edge-Aware Siamese Network (MAEANet) is built using a Siamese auto-encoder network intaking bi-temporal images, a multi-scale attention (MA) module based on the Convolutional Block Attention Module (CBAM) and the proposed Contour Channel Attention Module (CCAM) to enhance the differentiation of features and improve detection of false positive changes, and the edge-aware (EA) module to reduce confusion of boundaries of dense buildings. An ablation study (based on the MA and EA modules) showed that the combination together (MAEANet) yielded the best CD results with the highest F1-scores for both datasets. In addition, the model was experimented with two large building CD datasets and compared with four deep learning CD models, where the proposed MAEANet outperformed all 4 models in all accuracy metrics for both datasets. The study also included sensitivity tests on the scale factor for the hybrid loss (Weighted Binary Cross Entropy and Dice) and the optimal number and configuration of CBAM and CCAM for the MA module. The manuscript is well-structured and the reasoning behind the model architecture and its significance are presented in a systematic and clear manner. Also, the model shows superior CD performance with respect to the compared deep learning models, highlighting the significance of the authors’ contributions with ample novel concepts (ie. CCAM, EA, etc).

This is a strong manuscript with sound research results. Please also review the comments and suggestions below. Overall, the manuscript can be fine-tuned to address some minor tweaks with formatting and word choice. Also, please proofread the entire manuscript to make sure the usage of abbreviations and the grammar are consistent.

Please see the attached file for comments and suggestions. 

Comments for author File: Comments.pdf

Author Response

Dear Reviewer, 

Thanks for the useful comments and suggestions, please find the item-by-item response in the attached file. 

best regards, 
Xin Su

Author Response File: Author Response.pdf