This paper introduces BRIGHT, a novel and timely benchmark dataset for building damage assessment using multimodal high-resolution optical and SAR imagery. Covering 14 globally distributed disaster events, BRIGHT provides pixel-level damage annotations for over 384,000 buildings. The dataset is designed to facilitate AI-based disaster response research, particularly in challenging all-weather conditions. The authors also benchmark a suite of machine learning and deep learning models on multiple tasks. The authors provided detailed documents and descriptions, making the data, related source code, and pretrained weights of models easy to understand and use.

In summary, this is quite interesting and solid work. I’d like to recommend the acceptance of this work since it represents an important contribution to Earth observation and disaster response communities. Yet before acceptance, several clarifications and refinements are suggested.

1. The manuscript would benefit from deeper exploration of what the models learn from multimodal fusion. Specifically, what roles do optical images play in multimodal building damage assessment? Is it beyond just building footprint localization? On the other words, are the features extracted from optical imagery actively compared with SAR representations? Some discussion (e.g., based on CAMs in Fig. 7) is provided but can be more explicitly elaborated.
2. The manuscript makes extensive evaluation of supervised and unsupervised change detection models, but the conceptual and methodological relationship between building damage assessment and generic change detection remains unclear, which is largely implied rather than discussed. An explicit and clearer explanation would be great for readers who lack of related background.
3. Since UMCD methods underperform, consider including a random guessing baseline for reference. This would contextualize the difficulty of BRIGHT and help readers understand the performance floor under UMCD setup.
4. While Table 1 offers a comprehensive comparison of datasets, several datasets seem relevant and should be included to enhance its completeness, like CRASAR-U-DROIDs [arXiv:2407.17673] and Noto-Earthquake building damage dataset [10.5194/essd-2024-363].
5. The paper describes careful multimodal alignment but omits the software used, e.g., ENVI, ArcGIS, or QGIS. Please provide related details.
6. Appendix G includes important new experimental setups and evaluation methods for UMCD. However, too much content is composed together now. It is not easy for people to grasp information. Adding section subtitles could improve readability.
7. 8: Please specify in the figure or caption that the values represent average ± standard deviation across models.
8. 10: Add a note in the caption to clarify that each dot corresponds to performance on a single test event under cross-event transfer.
9. Typo in Table 7: “Object-based major voting” should be corrected to “Object-based majority voting”.
10. Clarify the meaning of “–” symbols in Table 11. Do they indicate missing data or inapplicability? This should be stated explicitly.
11. “ML” should be defined on its first use and consistently used thereafter instead of alternating with [machine learning].
12. Standardize currency formatting (e.g., USD vs. US$).
13. Define abbreviations such as IGN and GSI when first mentioned as data providers.
14. The format of references should be standardized. Some of these entries use abbreviations for journals, while others have full titles.

This manuscript introduces the BRIGHT dataset, which is a first open building damage assessment dataset with global coverage, multi-hazard scenarios, multimodal imagery (Optical and SAR), and sub-meter resolution. The paper systematically describes data collection, annotation, and quality control methods, and validates the dataset with multiple deep learning models, including cross-disaster transfer (zero-shot and one-shot), semi-supervised, and unsupervised approaches. The dataset demonstrates clear novelty and practical value, and it is of significant importance for advancing research and applications in disaster emergency response, remote sensing, and artificial intelligence. Generally, the paper is well structured, logically clear, with detailed results and strong value in terms of data sharing. Although the manuscript is rich in content, there are still details that require improvement, and I recommend appropriate revisions.

Major Comments

1. The explanation of annotation consistency and reliability remains insufficient. Although the authors state that the data annotations were obtained from multiple institutions such as Copernicus EMS, UNOSAT, and FEMA and then refined manually, there may be inconsistencies in how different institutions define “damaged” and “destroyed.” This could affect the consistency of annotations across disaster scenarios. It is therefore necessary to further elaborate on the process of unifying annotations, provide more detail on the manual refinement procedures.
2. The treatment of class imbalance is not sufficient. Figure 5(d) shows that intact buildings account for over 80%, while destroyed buildings account for less than 7%. This severe imbalance directly affects the accuracy of recognizing destroyed classes. Although the authors employed the Lovasz loss function to partially alleviate the issue, this is still not enough to solve the problem. Is this imbalance one of the reasons for the relatively low performance of the subsequent experimental results?
3. The discussion on cross-disaster generalization needs to be strengthened. Table 6 shows that in different disaster types, certain events perform particularly poorly. In particular, the mIoU values are the lowest for explosion and chemical accident events such as Bata-EP-2021 and Kyaukpyu-CC-2023, while earthquake events such as Morocco-EQ-2023 and Noto-EQ-2024 also remain highly challenging. This indicates that the models face significant difficulties in handling highly destructive, structurally complex, and spatially heterogeneous disaster scenarios. The authors should analyze these challenges in more depth, such as the heterogeneity and extreme local variations in explosion damage, the diversity of collapse patterns in earthquake events, and the limitations of SAR data in capturing fine-grained details. It is also recommended to provide typical error cases and compare model errors across disaster types to better illustrate the shortcomings in generalization.
4. The manuscript currently lacks a comparison with optical-only baselines, which is crucial to highlight the value of multimodal methods. Readers may question whether the inclusion of SAR brings significant benefits and whether the additional cost of multimodality is justified. To avoid such doubts, I suggest adding experiments with optical-only inputs and comparing them with Optical+SAR results. This would further emphasize the unique value of the BRIGHT dataset and provide stronger evidence for the necessity of multimodal fusion.
5. The discussion of limitations and future directions is insufficient. At present, the conclusion mainly emphasizes the dataset’s contributions, but it does not address its shortcomings in detail. It is suggested to include a separate subsection summarizing the limitations, such as the use of single-polarization SAR, the lack of time-series data, and the fact that most disaster events are concentrated after 2020.

Minor Comments

1. The description of study areas and disaster events is somewhat redundant.
2. The terminology for “one-shot” may not be accurate. The authors describe it as using “a small number of labeled samples,” which may be better defined as “few-shot.” Since these concepts are borrowed from previous work, it is suggested to cite the corresponding references.
3. At line 420, the authors state that SAR is not sensitive to fine structural changes. Would this limitation reduce the value of multimodality in certain scenarios?