the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 May 2026
Improved Dating of Landslides in Zimbabwe by Combining Satellite Multispectral and Synthetic Aperture Radar Observations
Abstract. Accurate dating of landslides is essential for understanding triggering mechanisms and improving hazard analysis, yet many inventories lack precise event timing. This study presents a two-step methodology for dating existing inventories using multi-sensor satellite data and automated change-point detection implemented with the Ruptures Python package. In Step 1, extended time series of Sentinel-2 optical NDVI and the Bare Soil Index are analysed to estimate the approximate dates of landslides. Step 2 refines these estimates using Sentinel-1 SAR VV backscatter data within a six-month window centred on the results from Step 1. The approach is tested using a landslide inventory from Zimbabwe associated with Storm Idai in March 2019. Using the results from Step 2, 52.8 % of the dataset is dated, with 84.6 % accuracy (events correctly dated during the triggering storm event) and an average precision of 12.8 days (the dates between which the event occurred). The results demonstrate that combining optical and SAR satellite observations with automated change-point detection provides an effective method for retroactively dating landslides. This approach enables inventories to be dated with minimal prior knowledge of event timing or geometry, while avoiding the need for large datasets and high-performance computing resources. The code is made available in Google Earth Engine, allowing for wide application.
- Preprint
(14862 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-2418', Katy Burrows, 27 May 2026
-
CC1: 'Comment on egusphere-2026-2418', Olivier Dewitte, 03 Jul 2026
Dear Joanna Moyes and co-authors,
I read your study on landslide dating in Zimbabwe with great interest. To further enrich the discussion, I would like to draw your attention to two recent studies that, I believe, are relevant.
- First, the work by Dille et al. (2026). This study investigates the same event as your paper, i.e. the landslides triggered by cyclone Idai. This work, that was published after your manuscript has already been submitted for review, provides a substantially more comprehensive landslide inventory. Nearly 12,000 landslide source areas (predominantly shallow) are mapped as polygons, together with a comparable number of runout areas. Overall, this inventory contains approximately ten times more mapped features than the inventory of Emberson et al (2022), which forms the basis of your analysis. I believe these differences, which may raise questions about the completeness and representativeness of the Emberson et al. (2022) inventory, could be acknowledged and briefly discussed in your manuscript.
      The inventory is available here: https://ihp-wins.unesco.org/dataset/landslide-susceptibility-exposure-assesment-chimanimani-chipinge
- Second, the work by Deijns et al. (2024). This study presents a semi-automatic approach for landslide detection based on times series of eight spectral indexes, including NDVI. The approach of Deijns et al. (2024) was calibrated and validated using rainfall triggered landslides in Africa as well and relies on Sentinel-2 imagery. In addition to mapping landslides, the approach enables the timing of landslide occurrence to the determined. The resulting inventory comprises nearly 4,000 mapped landslides and may therefore provide useful context for your discussion on landslide dating methodologies
I hope you find these references helpful.
Kind regards,
Olivier Dewitte
References
Deijns, A.A., Michéa, D., Déprez, A., Malet, J.P., Kervyn, F., Thiery, W. and Dewitte, O., 2024. A semi-supervised multi-temporal landslide and flash flood event detection methodology for unexplored regions using massive satellite image time series. ISPRS Journal of Photogrammetry and Remote Sensing, 215, pp.400-418.
Dille, A., Dewitte, O., Broeckx, J., Verbist, K., Sindiso Dube, A., Poesen, J. and Vanmaercke, M., 2026. Capturing the complete landslide–debris-rich flood continuum for accurate inventory, susceptibility and exposure mapping–lessons from Cyclone Idai. Natural Hazards and Earth System Sciences, 26(6), pp.2561-2577.
Emberson, R., Kirschbaum, D., Amatya, P., Tanyas, H. and Marc, O., 2021. Insights from the topographic characteristics of a large global catalog of rainfall-induced landslide event inventories. Natural hazards and Earth system sciences discussions, 2021, pp.1-33.
Citation: https://doi.org/10.5194/egusphere-2026-2418-CC1 -
RC2: 'Comment on egusphere-2026-2418', Anonymous Referee #2, 10 Jul 2026
This manuscript presents a practical and accessible workflow for retrospective landslide dating using Sentinel-2 multispectral indices (NDVI and BSI) combined with Sentinel-1 VV backscatter and automated change-point detection through the Ruptures package. The approach is attractive because it requires only landslide point locations and publicly available datasets within the Google Earth Engine/Google Colab ecosystem, making it potentially useful for researchers working with historical inventories that lack detailed geometries or event dates. The manuscript is generally well structured, and the workflow is clearly described. The study demonstrates promising results on a large inventory (1,319 landslides), and the comparison with previous studies is useful. However, before publication, several methodological issues require clarification. Most of my concerns relate to the generalizability of the workflow, the optimisation strategy, the statistical evaluation, and the interpretation of the reported performance. Overall, I believe the manuscript has merit and could become a useful contribution after major revision.
Major comments
1. The parameter optimisation appears to be performed on the validation dataset?
          Several methodological choices appear to have been selected using the same Zimbabwe inventory that is later used to evaluate performance. These include a 100 m background buffer, preferred Ruptures penalties, hierarchical ranking of change-point models, and a ±90-day SAR window, filtering based on physically plausible NDVI/BSI changes. Although these choices are individually reasonable, they collectively raise concerns regarding overfitting. The manuscript would be considerably strengthened if the authors clearly separated parameter calibration, parameter validation, and final model evaluation. If an independent validation dataset is unavailable, the authors should, at a minimum, explicitly discuss this limitation and quantify the sensitivity of the results to these parameter choices.
2. Sensitivity analysis may require more empirical testing
          Many important parameters appear to have been fixed after empirical testing, but the manuscript only briefly explains why. For example, why exactly 100 m?, why ±90 days rather than ±60 or ±120?, why L2 penalty = 1 for SAR?, and why this specific hierarchy of penalties? Although Figures 5 and 7 partially justify these decisions, a more systematic sensitivity analysis would greatly improve confidence in the methodology. Ideally, the supplementary material could include performance curves showing how accuracy, precision, and percentage of dated landslides vary across parameter space.
3. It would be good for the readers to have a definition of the evaluation metrics used in this study.
The manuscript uses terms such as accuracy, precision, correctly dated, overlap, and percentage dated throughout. However, these definitions are not always intuitive. For example, 84.6% "accuracy" is defined as the percentage of dated landslides for which the estimated temporal window overlaps the known event window. This differs substantially from the conventional statistical definition of classification accuracy. I recommend including a dedicated subsection that defines each evaluation metric mathematically. Similarly, "precision" is used to describe temporal uncertainty rather than predictive precision, which may confuse readers.
4. The authors shall consider performing an uncertainty analysis.
Every detected landslide date is ultimately treated as deterministic. However, uncertainty exists at multiple stages, viz., Sentinel revisit interval, cloud masking, rupture change-point selection, SAR smoothing, and parameter selection. The manuscript would benefit from a discussion of how these uncertainties propagate through the workflow. For example, are there confidence scores associated with individual detections? Could multiple candidate change points be retained rather than selecting only the first one?
5. Discussion of computational efficiency could be expanded
The manuscript reports that approximately 100 landslides can be processed in 45–60 minutes. This implies roughly 10–13 hours for the full inventory. Since one of the principal advantages of the proposed workflow is accessibility, readers would benefit from additional discussion regarding computational bottlenecks, opportunities for parallelisation, use of GEE map/reduce operations, and scalability to inventories containing tens of thousands of landslides.
6. Comparison with previous studies shall be more balanced
Table 1 compares numerous studies using different datasets, landslide types, spatial resolutions, inventory quality, vegetation conditions, and validation strategies. Consequently, direct comparison of percentages dated or reported accuracy may not always be meaningful. The discussion should acknowledge these differences more explicitly to avoid implying superiority where experimental conditions differ substantially.
Minor Comments
- The manuscript repeatedly states that the workflow "requires only point locations." Strictly speaking, this is true for the execution of the workflow, but the optimisation and validation were performed using polygon inventories converted to centroids. This distinction should be stated more explicitly.
- The discussion could better explain why NDVI and BSI were selected instead of other vegetation or disturbance indices (e.g., NBR, EVI, NDMI). Readers may wonder whether these alternatives were evaluated.
- The rationale for using only VV polarisation deserves slightly more explanation. Although Figure 7 shows better performance than VH, quantitative statistics comparing the two channels would improve transparency.
- Figure 5 is somewhat difficult to interpret because multiple penalty values and methods are plotted together. A clearer legend or separate panels would improve readability.
- Several figures display only one representative landslide. It would be valuable to include additional examples illustrating successful and failed detections, as well as noisy optical and SAR time series. These examples would help readers understand the robustness of the workflow.
- The manuscript frequently uses the phrase "dated correctly." Since the true date is only constrained to a temporal window (15–20 March), wording such as "correctly assigned to the known event window" would be more precise.
- The authors mention false positives using randomly selected non-landslide points. This is an interesting experiment and could be expanded slightly. For example, what land-cover classes generated the most false positives? Could these be filtered automatically?
- The Conclusions occasionally repeat numerical results already discussed extensively. This section could instead focus more on the broader implications and future applications of the workflow.
- There are occasional inconsistencies in capitalisation (e.g., "Step 1", "step 1"). Some figure captions are overly detailed and could be shortened. Ensure all abbreviations (AOI, BSI, VV, VH, L1, L2) are defined at first appearance.
- Carefully proofread for minor grammatical issues and repeated wording.
Citation: https://doi.org/10.5194/egusphere-2026-2418-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 229 | 76 | 19 | 324 | 16 | 15 |
- HTML: 229
- PDF: 76
- XML: 19
- Total: 324
- BibTeX: 16
- EndNote: 15
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Please see the attached document