the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 17 Feb 2026
Climate change increases landslide susceptibility in Aotearoa New Zealand: Development and application of a national-scale model using machine learning
Abstract. Rainfall-induced landslides (RILs) pose a major hazard to infrastructure, ecosystems, and communities across Aotearoa New Zealand, with events such as Cyclone Gabrielle underscoring the potential scale of their impacts. In this study, we develop a relatively high-resolution national-scale RIL susceptibility model that includes both conditioning and triggering variables and use it to assess the impacts of climate change on RIL susceptibility. The model utilises machine learning (ML) (gradient boosted decision trees) to predict RIL susceptibility in response to extreme rainfall events under current and future climate scenarios at 25 m spatial resolution. We use a training dataset of observed landslides triggered by Cyclone Gabrielle in the Hawke's Bay and Gisborne/Tairāwhiti regions. Predictor variables include topographic, geologic, and environmental factors, with rainfall intensity serving as the primary trigger. Model performance is evaluated using Shapley additive explanations (SHAP) analysis, alongside standard error metrics, achieving a receiver operating characteristic area under the curve (ROC-AUC) of 0.94. We then apply the model nationally to estimate RIL susceptibility under six current and 24 future storm scenarios based on NIWA’s high-intensity rainfall design system (HIRDS) datasets and modelled temperature changes under different shared socioeconomic pathways (SSPs). Results show a substantial increase in RIL susceptibility under warmer climate futures, with susceptibility increasing disproportionately to rainfall increase. Forest cover is found to play an important role in mitigating susceptibility. This work presents a robust framework for national-scale RIL susceptibility assessment under specific storm scenarios and provides a national-scale dataset suitable to support climate-resilient land use planning and nature-based mitigation strategies.
- Preprint
(2883 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-768', Anonymous Referee #1, 16 Apr 2026
-
RC2: 'Comment on egusphere-2026-768', Anonymous Referee #2, 14 Jun 2026
General comments
The manuscript focuses on changes in national-scale rainfall-induced landslide susceptibility in New Zealand under the background of climate change. The topic has practical significance and is also within the scope of NHESS regarding natural hazards, climate change impacts, and risk assessment. Based on landslide data triggered by Cyclone Gabrielle, the author developed a LightGBM model and combined topographic, geological, soil, vegetation, and rainfall factors to predict landslide susceptibility under current and future SSP scenarios. Overall, the research idea is relatively clear, and the framework also has certain application value. However, the current manuscript still has several key methodological issues. The landslide inventory is mainly constructed based on NDVI-derived bare-ground change and a slope threshold, which introduces certain uncertainty in sample classification. The model training mainly relies on samples from the single Cyclone Gabrielle event and a limited region, but the model is then extended to the national scale and to different future storm scenarios, while its generalization ability has not been sufficiently validated. Important influencing factors such as antecedent soil moisture and aspect were not included. The resampling of HIRDS rainfall data from 2000 m to 25 m may also introduce scale mismatch and spatial bias. In addition, although the ROC-AUC is high, the PR-AUC for the landslide class is relatively low, and at the 5% threshold, false positives are clearly more numerous than true positives, indicating that the model has a clear tendency toward over-prediction. Therefore, I believe that the current results are not sufficient to support the main conclusions of the manuscript regarding changes in landslide susceptibility at the national scale and under future climate scenarios. The author is advised to consider resubmission only after further improving the landslide inventory, adding validation using independent events or independent regions, conducting sensitivity analyses of thresholds and prior probabilities, and fully discussing the uncertainty of future rainfall data. For the above reasons, I recommend that the current manuscript not be accepted.
Specific comments
- The rationale for the construction of the landslide inventory and the selection of thresholds needs to be further explained. In the manuscript, Sentinel-2 images before and after Cyclone Gabrielle are used to identify bare-ground changes through NDVI changes, and then a slope >10° rule is applied to distinguish landslides from flood deposits and other areas. This approach is reasonable to some extent, but NDVI essentially identifies bare-ground change rather than landslides themselves. At the same time, the slope >10° threshold is an empirical threshold, which may affect which areas are ultimately defined as landslide samples. The author is advised to explain the basis for selecting the NDVI threshold and the slope threshold, and to further discuss whether different slope thresholds would lead to obvious changes in the number of landslide samples, their spatial distribution, and the model results.
- The author combined the landslide data identified in this study with the landslide inventory from Dragonfly Data Science, and only used pixels identified as landslides in both datasets as positive samples. Areas identified by only one of the two datasets were classified as “possible landslides” and excluded from both training and testing. This approach can reduce misclassification, but it may also make the training samples more biased toward relatively obvious and larger landslides, causing the model to be more capable of identifying large areas of bare-ground change rather than necessarily representing all rainfall-induced landslides. The author is advised to further explain whether this treatment affects the completeness of the landslide samples, and whether excluding the “possible landslide” areas changes the spatial representativeness of the non-landslide samples.
- Regarding the representativeness of the training area and the issue of national-scale extrapolation, the author is advised to provide further justification. The model is mainly trained using landslide samples from the area affected by Cyclone Gabrielle, but it is ultimately applied to the whole of New Zealand. Considering that different regions may vary in terms of topography, geology, land cover, and rainfall characteristics, whether a model trained only on this event and this region is sufficient to support national-scale prediction needs to be more fully explained.
- The model training samples come from the single event of Cyclone Gabrielle. This event has its own rainfall path, duration, storm direction, antecedent wetness conditions, and regional spatial characteristics. Therefore, whether the model can be generalized to other rainfall events still needs further discussion. If there is no validation using other independent rainfall events, the author is advised to be more cautious when describing the applicability and generalization ability of the model.
- Rainfall-induced landslides are usually affected not only by event rainfall, but also closely related to pre-storm soil moisture conditions. The author is advised to further discuss the possible impacts of the absence of antecedent soil moisture, and to consider whether antecedent rainfall indices, soil moisture reanalysis data, or other data could be used for supplementary testing or sensitivity analysis.
- The author did not include the aspect variable, on the grounds that the influence of aspect on landslides may depend on storm direction, while future storm direction is unknown. This explanation is understandable, but for the specific training event of Cyclone Gabrielle, aspect and storm direction may still have influenced the landslide distribution. The fact that storm direction cannot be determined in future scenarios does not mean that aspect is completely unimportant in the current event-based modelling. The author is advised to further explain the possible influence of excluding the aspect variable on model interpretation and prediction results.
- Regarding the accuracy and applicability of the future precipitation data itself, the author is advised to provide further explanation. Precipitation has significant spatial heterogeneity, and whether HIRDS data are reliable in representing the spatial differences in future precipitation needs to be clarified. In addition, the original resolution is 2000 m, and it is resampled to 25 m; this process may further introduce spatial bias. The author is advised to discuss whether different resampling methods, such as nearest neighbour, cubic convolution, and inverse distance weighting, would have a significant impact on the landslide susceptibility prediction results.
- The model results show an increase in landslide susceptibility under future warming scenarios, and this direction is reasonable. However, the magnitude of the increase may be affected by factors such as the probability threshold and the prior probability setting. The author is advised to further explain whether the magnitude of the future increase in landslide susceptibility remains stable under different thresholds or different prior probabilities.
- Although the ROC-AUC is high, the PR-AUC for the landslide class is relatively low, and at the 5% threshold, false positives far outnumber true positives. The author is advised to make this point clearer in the interpretation of the results, and to further discuss the possible effects of this over-prediction on the landslide susceptibility maps, the estimated affected area, and the exposure assessment of buildings and roads.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,295 | 675 | 75 | 2,045 | 50 | 114 |
- HTML: 1,295
- PDF: 675
- XML: 75
- Total: 2,045
- BibTeX: 50
- EndNote: 114
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript focuses on the development of a national-scale rainfall-induced landslide model for New Zealand and the projected changes in landslide susceptibility under future climate scenarios. The topic is of clear practical relevance. The authors attempt to incorporate triggering rainfall into the model and further combine the modelling results with future climate scenarios and exposure analyses for roads and buildings, thereby building a national-scale application framework. From a methodological perspective, the manuscript include landslide sample extraction, predictor selection, LightGBM modelling, SHAP interpretation, and scenario-based projection. The amount of work is substantial, and the results have reference value. However, there are several important issues still require further clarification.
1. The training data are based on landslides triggered by Cyclone Gabrielle in Hawke’s Bay and Gisborne/Tairāwhiti. The authors also acknowledge that the South Island, especially the Southern Alps, differs substantially from the training area in both lithology and dominant slope processes. Pages 21-22 further state that the South Island is dominated by harder metamorphic rocks, and that the Southern Alps commonly feature scree/talus slopes, rock avalanches, and large deep-seated landslides, which are not equal to the shallow rainfall-induced landslides that are more typical of the North Island. The manuscript also notes that “the interpretation of RIL susceptibility maps for the Southern Alps should be undertaken cautiously.” It is suggested that the authors either state the applicable scope more accurately in the title and main conclusions, or add cross-regional validation.
2. The authors identify bare-ground change using a ΔNDVI threshold derived from pre- and post-event multi-temporal Sentinel-2 imagery, and then filter landslide pixels using a slope threshold of >10°. As explicitly stated on page 5, this approach “includes both the initiation point and run-out zone.” This means that the model is not learning purely the conditions of landslide initiation, but rather the spatial pattern of terrain affected after the event. On page 20, the authors also state that the model “does not separate the RIL initiation point from the debris run-out zone.” This has direct implications for the physical interpretation of predictors such as slope, curvature, TWI, and soil depth, because the geomorphic characteristics of runout zones differ from those of initiation zones. The authors should clarify more explicitly whether the model output is closer to initiation susceptibility, event-affected area probability, or a mixed footprint probability.
3. The authors use slope >10° to distinguish landslide-related bare-ground change from non-landslide areas. On page 13, in the SHAP interpretation, the manuscript itself states that the lower susceptibility below 10° “may also be a consequence of the RIL mapping exercise.” This indicates that part of the slope effect is directly embedded in the sample construction procedure, rather than being entirely learned by the model from independent observations of the process. Since slope is identified as the second most important predictor, this issue should be explained more carefully in the Results and Discussion.
4.The authors first hold out 20% of the landslide polygons, then randomly sample non-landslide pixels from the full region, and use a 10 m exclusion buffer and a 72 m buffer to reduce spatial leakage. Even so, the training and testing data still come from the same Cyclone Gabrielle event and the same broader regional context. In addition, the 10-fold cross-validation used in the manuscript is random cross-validation and does not explicitly account for spatial autocorrelation. It is suggested that the authors further evaluate the model using spatial cross-validation.
5. The future scenarios are derived from HIRDS rainfall data at 2000 m resolution, statistically downscaled to 25 m, and then combined with the change factors in Table 2 and the national mean temperature changes in Table 3 to construct the SSP2-4.5 and SSP3-7.0 scenarios. On the one hand, the future intensification signal is based on national mean warming rather than finer-scale regional climate differences. On the other hand, the manuscript does not systematically report the uncertainty range across the six climate models, but instead directly uses the national mean warming signal. The resulting outputs can be considered as national-scale scenario projections, but they may not be sufficient to support more detailed local-scale interpretations. It is suggested that the authors add corresponding statements or uncertainty discussion in the Abstract, Discussion, and Conclusion.
6.The presentation of the SSP scenarios should be made consistent throughout the manuscript. It would be better to use SSP2-4.5 and SSP3-7.0 consistently, rather than 2-4.5 and 3-7.0.