Rapid Probabilistic Inundation Mapping Using Local Thresholds and Sentinel-1 SAR Data on Google Earth Engine
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe main research content of this paper is the proposal of a rapid probabilistic flood mapping method. This method combines local thresholds derived from Sentinel-1 Synthetic Aperture Radar (SAR) images and land cover data to estimate the probability of surface water, utilizing Google Earth Engine for fast probabilistic flood mapping. The paper also validates the effectiveness and efficiency of the method by testing it on various flood events across five continents, demonstrating its reference value. However, there are some issues that the authors need to consider and address, as detailed below:
1. The term "VV" mentioned in the abstract is unclear upon first reading—its meaning should be clarified.
2. The claim in the introduction that binary classification ignores detection uncertainty is inappropriate. Modern binary classification in imaging also generates probabilities, provides thresholds, and can even perform dynamic probability calibration and multi-dimensional uncertainty quantification to improve classification accuracy.
3. The sample sizes for different land cover types vary significantly, and there may also be substantial differences between datasets from different time periods. This not only affects model training performance but also raises concerns about the model's applicability to new data. How do the authors account for these discrepancies and improve the model?
4. The authors selected only 11 flood events for analysis, which is a relatively small sample size. Do these events cover regions with diverse geographical environments? How is the model's generalization capability ensured to avoid overfitting?
5. Although the Google Earth Engine platform offers fast processing speeds, delays can occur in data acquisition and preprocessing. How do the authors ensure real-time processing?
6. While the local thresholding method may perform well within specific regions and existing samples, how is its applicability ensured across different climatic zones?
Author Response
Please see the attachment. Thank you!
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors
- The author claimed that the code is publicly shared, but I couldn’t find a link.
- Gamma distributions are used to model SAR backscatter, which is reasonable for some land cover types, but the selection of Gamma distributions is not always optimal across all environments. Gaussian distributions could be more appropriate for certain land types. A sensitivity analysis on distribution choice would make the model more robust.
- The method’s accuracy varies significantly by land cover type (e.g., poor performance in bare and wetland regions). The paper acknowledges this but could explore solutions, such as integrating auxiliary datasets (e.g., LiDAR, HAND) or refining land cover classifications.
- Sentinel-1’s revisit time (~6 days) may miss short-duration floods. The paper could discuss how temporal gaps might be mitigated, e.g., by fusing optical data (where available) or using higher-frequency SAR missions (e.g., NISAR).
- Copernicus land cover data at 100 m resolution might not capture fine-scale heterogeneity. Higher-resolution land cover products (e.g., 10-m ESA WorldCover) could improve local thresholding. The author should discuss it.
- The Sen1Floods11 dataset, while comprehensive, may not represent all flood scenarios (e.g., urban flash floods). Including more diverse events (e.g., snowmelt floods, coastal surges) would strengthen validation.
- The hand-labelled water masks rely on Sentinel-2, which is unavailable during cloud cover. Cross-validation with independent ground truth data (e.g., UAVS, gauges) would bolster reliability.
- While GEE enables scalability, the paper does not quantify processing times for large regions. Benchmarking against other GEE-based flood algorithms (e.g., [58]) would highlight operational advantages.
- VV polarization outperforms VH, but the rationale (e.g., sensitivity to surface roughness) is not deeply explored. A discussion on polarization selection criteria (e.g., urban vs. vegetated areas) would be valuable.
- The method does not explicitly distinguish floodwater from permanent water bodies. Incorporating temporal baselines (e.g., median pre-flood images) or water occurrence datasets (e.g., JRC Global Surface Water) could enhance specificity.
- The impact of land cover misclassification (e.g., CGLS-LC100’s ~80% accuracy) on thresholding is not quantified. Error propagation analysis would help assess confidence in probabilistic outputs
- To improve robustness, combine Sentinel-1 with Sentinel-2 (for cloud-free periods), altimetry data (for water levels), or social media feeds (for ground validation).
- Implement machine learning (e.g., Random Forests) to dynamically adjust thresholds based on seasonal backscatter variations or real-time weather data.
- Develop interactive tools to visualize probabilistic outputs alongside uncertainty metrics (e.g., confidence intervals) for decision-makers.
- To evaluate its limits, Test the method in extreme environments (e.g., ice-covered floods, arid regions).
Author Response
Please see the attachment. Thank you!
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis is a very interesting study in the field of flood detection using remote sensing technology. Below are some comments and suggestions to help clarify and improve the manuscript:
-
Please provide a table listing all sample locations used in your study (including the name and area of each location), as referenced in Figure 1. Why were these specific locations selected as showcase areas?
-
The legend in Figure 4 is too small. Please enlarge it to ensure readability.
-
A cross-validation of the flood mapping should be conducted for some of selected areas (Figure 1). This would help to highlight the effectiveness of the proposed method compared to existing methods discussed in previous studies (see lines 64–75).
-
The Conclusions section is currently quite brief. Please expand it by including more bullet points summarizing the main contributions and highlighting the key results of this study.
Author Response
Please see the attachment. Thank you!
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors provided detailed explanations and revisions to the questions I raised, and added limitations to the discussion section, which dispelled my doubts about this study. Therefore, I believe that this study can be published in this journal.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe author has addressed all comments, and the manuscript can be accepted for publication.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe revised manuscript is qualified for publication.