the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 10 Jan 2023
More than one landslide per road kilometer – surveying and modelling mass movements along the Rishikesh-Joshimath (NH-7) highway, Uttarakhand, India
Ravi Kumar Guntu
Alexander Plakias
Igo Silva de Almeida
Wolfgang Schwanghart
Abstract. The rapidly expanding Himalayan road network connects rural mountainous regions. However, the fragility of the landscape and poor road construction practices lead to frequent mass movements along-side roads. In this study, we investigate fully or partially road-blocking landslides along the National Highway (NH-) 7 in Uttarakhand, India, between Rishikesh and Joshimath. Based on an inventory of > 300 landslides along the ~250 km long corridor following exceptionally high rainfall in October and September 2022, we identify the main controls on the spatial occurrence of mass-movement events. Our analysis and modelling approach conceptualizes landslides as network-attached spatial point pattern. We evaluate different gridded rainfall products and infer the controls on landslide occurrence using Bayesian analysis of an inhomogeneous Poisson process model. Our results reveal that slope, rainfall amounts, and lithology are the main environmental controls on landslide occurrence. The individual effects of aggregated lithozones is consistent with previous assessments of landslide susceptibilities of rock types in the Himalayas. Our model spatially predicts landslide occurrences and can be adapted for other rainfall scenarios, and thus has potential applications for efficiently allocating efforts for road maintenance. To this end, our results highlight the vulnerability of the Himalayan road network to landslides. Climate change and increasing exposure along this pilgrimage route will likely exacerbate landslide risk along the NH-7 in the future.
- Preprint
(2343 KB) -
Supplement
(139719 KB) - BibTeX
- EndNote
Jürgen Mey et al.
Status: closed
-
RC1: 'Comment on nhess-2022-295', Anonymous Referee #1, 08 Feb 2023
This manuscript takes the Rishikesh Joshimath (NH-7) highway in North Akandu, India, as the research area or line. Based on field survey, after the rainfall in September and October 2022, the location information of more than 300 landslides was obtained along the road line. It is helpful to understand the mechanism analysis and spatial distribution law analysis of rainfall landslides in this area, and landslide disaster prevention and mitigation. And this manuscript has high efficiency and good timeliness. However, there are some major defects in this manuscript. The specific suggestions are as follows: (1) The means of data acquisition in this manuscript are very limited, and it is far from enough to rely only on field survey. This can only be limited to some targets, and cannot reflect the overall landslide disaster situation. It is suggested that the author consider introducing multi-temporal satellite images. Moreover, relying only on the field survey, it can not be objectively judged that there is an inevitable relationship between these landslides and the rainfall in 2022. (2) The results obtained from incomplete data are often inconsistent with the actual situation. The statistical analysis results in this manuscript are affected by the limitations of landslide investigation technology. It is undoubtedly very difficult to reach the conclusion that "slope, rainfall amounts, and lithology are the main environmental controls on landslide occurrence", and lack of reliable model support. (3) Because this model is too local regional and so limitation of landslide data, it is difficult to popularize and apply it in other regions, and it is also difficult to obtain more general results. To sum up, due to the inadequacy of the underlying data and methods, I suggest reject the manuscript.
Citation: https://doi.org/10.5194/nhess-2022-295-RC1 -
AC1: 'Reply on RC1', Jürgen Mey, 08 Feb 2023
We thank the reviewer for the comments on our manuscript. We summarize the following points raised by the reviewer (R) and provide our replies (A):
R1: It is far from enough to rely only on field survey for data acqusition as it cannot reflect the overall disaster situation. The authors should consider using multi-temporal satellite images.
A1: We thank the reviewer for the suggestion to use multi-temporal imagery. In fact, we checked each landslide location using imagery made available by Google Earth using the latest and historic imagery (L. 112). We classified each location as new landslide, visible before the Sep-Oct 2022 rainfall anomaly and reactivated landslide. We note that detecting landslides from imagery was not always straightforward since landslide scars are difficult to distinguish from unvegetated engineered slopes and road widening (L. 115).
In fact, we emphasize that our mapping approach enabled us to detect the often small landslides lining the road. Moreover, we are confident that our landslide inventory of (partially) road blocking landslides is near to complete and thus reflects the overall disaster situation quite well.
In a revised version of the manuscript, we will document the visual interpretation of the imagery and its results more carefully.
R2: It can not be objectively judged that there is an inevitable relationship between these landslides and the rainfall in 2022.
A2: We raised this point in the text (L. 185):
"Yet not all field-mapped landslide occurrences can be attributed to the anomalously high rainfall period during September and October 2022. Visually inspecting the locations using Google Earth reveals that 21.4 % of the recorded landslides with road blockages existed before (Figure 1). 17.8 % of the landslides were most likely reactivated by the excessive rainfall because they could not be identified to be road-blocking before the rainfall period. Most landslides (60.8 %) were not identifiable as such in the Google Earth imagery available for several dates before and including March 2022."
We ran the analysis with a subset of 60.8% and the results are very similar to those obtained when taking all landslides into account. In a revised version, we will document this analysis.
R3. Incomplete data leads to results inconsistent with the actual situation. The statistical results are affected by limitations of the landslide investigation technology, and the conclusion that "slope, rainfall amounts, and lithology are the main environmental controls on landslide occurrence" are not substantiated by the model.
A3: We disagree with the reviewer. It is the aim of statistical modelling to infer models/test hypotheses based on incomplete data. We regard the actual pattern of landslides as a point pattern and infer the environmental controls using established statistical models. Our inference is complemented by a careful Bayesian uncertainty analysis (Fig. 4, 6 and 7). Based on our model, slope, rainfall amounts and lithology explain the observed landslide patterns very well. Of course, statistical models enable us to infer correlation, not causation. The mechanistic interpretation of the correlations are subject of the discussion.
R4: The model does not provide more general results and cannot be applied to other regions.
A4: This comment addresses nearly any field-based study. Whether our model can or cannot be applied to other regions is elusive and remains to be tested. The consistency of the individual effects of lithozones with previous assessments of landslide susceptibilities of different Himalayan lithologies suggests, however, that our model is applicable beyond the studied extent of the NH-7.
With kind regards,
Jürgen Mey on behalf of all coauthors.
Citation: https://doi.org/10.5194/nhess-2022-295-AC1
-
AC1: 'Reply on RC1', Jürgen Mey, 08 Feb 2023
-
RC2: 'Comment on nhess-2022-295', Anonymous Referee #2, 30 Mar 2023
The manuscript deals with an interesting and largely discussed (far from being solved though) topic in the landslide susceptibility scientific community, i.e. the inter-relationship between linear communication infrastructures and landslides. Specifically, this work aims to carry out a landslide susceptibility analysis for a ~250 km long stretch of the National Highway 7 (NH-7), India.
The relevance of the addressed topic is well justified and supported by adequate references. The authors propose an innovative perspective by considering the landslides found along the road as network-attached spatial point pattern inspired by the work of Baddeley et al. (2021), which, from my point of view, is a potential line of research.
However, I believe that the manuscript shows some important issues that make it not suitable for publication in its current version. The objectives of the research are unclear, the applied methods are not explained with enough details and the structure of the information provided should be improved.
In the following lines I’ll try to explain the main weak points of the manuscript and how they could be improved in my opinion:
(1) The authors dedicated a big effort conducting a landslide inventory that collects more than 300 failures. I acknowledge the big value of such a database as a proxy of the exposure of the road against landslides. And in this sens, I do believe that the road and its exposure to small landslides could be modeled considering the morphological and environmental features throughout its margins. But, the authors focus the analysis to only road side occurring landslides stating that they "only mapped landslides with runouts affecting the road" (Line 109). So, the dataset is restricted to those observations that strictly has affected the road, limiting the representativeness of the inventory. In my opinion, with this dataset it is not possible to model the occurrence of landslides correctly, since it doesn’t consider the full picture of the landslide occurrences. I would suggest to change the perspective of the research and reformulate it as a road exposure research rather than landslide susceptibility study.
(2) The consideration of dynamic environmental maps, such as the cumulated rainfall, as one of the possible controlling factors is conceptually correct and its effect in the landslide occurrence is justified. However, these kind of inputs should be used with extreme care, and the suitability of rainfall products for spatial modelling have to go always very well supported. As an example, a cumulated rainfall map of 100 km of resolution is not going to be able to represent the rainfall variability within a study area of 250 km². In this work the rainfall distribution of that raining event has been represented by several cumulated precipitation maps, which the authors acknowledge to be considerable different among them. The authors select CHIRPS v2 based on the better performance of the model when this map is used as input, and there are no evidences provided that this map really represents the actual rainfall pattern in this area. In my opinion, the statistical results support the idea that the landslides in the inventory are controlled by the rainfall pattern displayed by CHIRPS v2; but, what we are not certain at all is wether this pattern is what has really happened or not, and if the original resolution of this map represents fairly the rainfall variability at the scale of the studied area. I suggest to include a complete section that deals with this issue, where the authors can justify very clearly why they use each product. This should include the assessment of the distortion of the original product, at its original resolution, with respect to the resampled one.
Another crucial aspect concerning the rainfall variables is the resampling method applied to convert the original rainfall products (in 0.05ºx0.05º to 0.25ºx0.25º of resolution) to the working environment of 30x30 meters of resolution. There is no information about that in the manuscript.
(3) I agree with the authors about the relevance of the slope gradient as a key controlling factor. But looking to the pictures provided in Fig. 2, it seems that the types of landslides in the inventory were triggered in the road cuts, just at a few meters of distance from the road. So, maybe the spatial resolution of 30x30 is not accurate enough to represent in a fair way the slope gradient at such a short distance. In addition, the followed approach to obtain and assign the slope value to the road pixel is not clear (Lines 144-146): Why is the buffer around the road set up to 210 m? What do you mean with the “nearest pixel”?
I suggest to grow up to a finer resolution DEM. CARTOSAT-1-derived DEM with 10x10 m of resolutions seems to be a good option (https://link.springer.com/article/10.1007/s40808-019-00694-9).
(4) The model validation approach should be described with more details. Are trained models validated with independent dataset? If yes, which criterion has been used to split the training and validation?
Since the collected data are exclusively related to one raining event, i.e., the monsoon season of 2022, I would try to feed the inventory with more landslides related to other rainfall events and use them for validation. Also, other roads that were affected by landslides for the same rainfall event could be appropriate to validate the model. And I acknowledge that the collection of this type of data is not straightforward, so if it’s not possible, another option could be to split the road stretch in several portions and apply a spatial validation approach.
To finish I’d like to recognise the big work that has been done to conduct the research presented in this manuscript and I find the proposed overall methodology very interesting, which probably can provide with promising results soon. Therefore, I encourage the authors to consider the comments in this revision so they can help to improve the work done. In addition, I briefly list some other minor suggestions.
- State very clearly the objectives of the research in the introduction section
- Provide the dates of the surveyed Google Earth images
- Provide the exact period that the cumulated rainfall maps represent
- When you talk about specific tools, such as TopoToolbox, specify in which package it works
- If you resample raster maps, specify the method used
- Add references about Akaike Information Criterion, the Receiver-Operating Characteristics or the Area under the Curve
- In Fig. 1 add Uttarakhand region in the location box
- In Fig. 3 add a scale to the maps
- I would also show the map of the resulting model so the reader can see, visually, the matching of the modelled susceptibility of the road and the landslide occurrences
Citation: https://doi.org/10.5194/nhess-2022-295-RC2 -
AC2: 'Reply on RC2', Jürgen Mey, 04 Apr 2023
A: We thank the reviewer (R) for helpful and constructive comments. Below, we provide replies to the comments and detail how we will address them in a revised version of the manuscript.
_________________
R: The manuscript deals with an interesting and largely discussed (far from being solved though) topic in the landslide susceptibility scientific community, i.e. the inter-relationship between linear communication infrastructures and landslides. Specifically, this work aims to carry out a landslide susceptibility analysis for a ~250 km long stretch of the National Highway 7 (NH-7), India.
The relevance of the addressed topic is well justified and supported by adequate references. The authors propose an innovative perspective by considering the landslides found along the road as network-attached spatial point pattern inspired by the work of Baddeley et al. (2021), which, from my point of view, is a potential line of research.
However, I believe that the manuscript shows some important issues that make it not suitable for publication in its current version. The objectives of the research are unclear, the applied methods are not explained with enough details and the structure of the information provided should be improved.
In the following lines I’ll try to explain the main weak points of the manuscript and how they could be improved in my opinion:
(1) The authors dedicated a big effort conducting a landslide inventory that collects more than 300 failures. I acknowledge the big value of such a database as a proxy of the exposure of the road against landslides. And in this sens, I do believe that the road and its exposure to small landslides could be modeled considering the morphological and environmental features throughout its margins. But, the authors focus the analysis to only road side occurring landslides stating that they "only mapped landslides with runouts affecting the road" (Line 109). So, the dataset is restricted to those observations that strictly has affected the road, limiting the representativeness of the inventory. In my opinion, with this dataset it is not possible to model the occurrence of landslides correctly, since it doesn’t consider the full picture of the landslide occurrences. I would suggest to change the perspective of the research and reformulate it as a road exposure research rather than landslide susceptibility study.
A: We agree with the referee that limiting the inventory to road blocking landslides may abate its representativeness for roadside landslides, in general. We emphasize, however, that we used this mapping criterion to cope with the overwhelming amount of landslides and to be able to take into account the landslides that detached most recently. Also, construction works for road-widening often stripped off vegetation so that differentiating between actual, non-blocking landslides and excavated slopes was not straightforward. In a revised version of the manuscript, we will provide more details on the motivation behind including road-blocking landslides only. Also, we like the suggestion to focus on road exposure rather than landslide susceptibility and reframe the study accordingly.
_________________
R: (2) The consideration of dynamic environmental maps, such as the cumulated rainfall, as one of the possible controlling factors is conceptually correct and its effect in the landslide occurrence is justified. However, these kind of inputs should be used with extreme care, and the suitability of rainfall products for spatial modelling have to go always very well supported. As an example, a cumulated rainfall map of 100 km of resolution is not going to be able to represent the rainfall variability within a study area of 250 km². In this work the rainfall distribution of that raining event has been represented by several cumulated precipitation maps, which the authors acknowledge to be considerable different among them. The authors select CHIRPS v2 based on the better performance of the model when this map is used as input, and there are no evidences provided that this map really represents the actual rainfall pattern in this area. In my opinion, the statistical results support the idea that the landslides in the inventory are controlled by the rainfall pattern displayed by CHIRPS v2; but, what we are not certain at all is wether this pattern is what has really happened or not, and if the original resolution of this map represents fairly the rainfall variability at the scale of the studied area. I suggest to include a complete section that deals with this issue, where the authors can justify very clearly why they use each product. This should include the assessment of the distortion of the original product, at its original resolution, with respect to the resampled one.
Another crucial aspect concerning the rainfall variables is the resampling method applied to convert the original rainfall products (in 0.05ºx0.05º to 0.25ºx0.25º of resolution) to the working environment of 30x30 meters of resolution. There is no information about that in the manuscript.
A: We agree with the referee that spatial rainfall variability is one of the weak points in our study. Clearly, we are limited by a lack of independent validation data which are either unavailable due to restricted accessibility or because they are already included in the development of the rainfall product. The strong difference between the rainfall products also highlights the apparent difficulties in predicting/hindcasting rainfall distributions in these steep mountain areas. As the referee notes, one could use the landslide data to infer the best rainfall product. However, this approach comes with the problem of circular argumentation. Gladly, and as pointed out by us in the manusript, the choice of rainfall distribution affects the overall predictability of road-blocking slope failures, but it doesn't affect the other parameters (slope and lithology). In a revised version of the manuscript, we plan to document better how we dealt with the rainfall data and provide more details on the bilinear resampling technique used to downsample the grids.
_________________
R: (3) I agree with the authors about the relevance of the slope gradient as a key controlling factor. But looking to the pictures provided in Fig. 2, it seems that the types of landslides in the inventory were triggered in the road cuts, just at a few meters of distance from the road. So, maybe the spatial resolution of 30x30 is not accurate enough to represent in a fair way the slope gradient at such a short distance. In addition, the followed approach to obtain and assign the slope value to the road pixel is not clear (Lines 144-146): Why is the buffer around the road set up to 210 m? What do you mean with the “nearest pixel”?
I suggest to grow up to a finer resolution DEM. CARTOSAT-1-derived DEM with 10x10 m of resolutions seems to be a good option (https://link.springer.com/article/10.1007/s40808-019-00694-9).
A: Yes, indeed the spatial resolution of the DEM may be too coarse to study phenomena occurring at finer spatial scales. Unfortunately, photogrammetric DEMs based on CARTOSAT imagery are available only at 30 m resolution and their quality is comparable to other globally available DEMs. We chose 210 m (7 x spatial resolution of the COP30 DEM) to have sufficient pixels to obtain a robust estimates of the slope which can fluctuate strongly in steep terrain. Our approach pixellates the road, and the model requires that the slope values are mapped to the road pixels. Thus, we identified the pixels closest and upslope to the road, and then aggregated slope values measured in up to 7 pixel distance to the nearest road pixels. In a revised version of the manuscript, we will provide more details about this approach.
_________________
R: (4) The model validation approach should be described with more details. Are trained models validated with independent dataset? If yes, which criterion has been used to split the training and validation?
Since the collected data are exclusively related to one raining event, i.e., the monsoon season of 2022, I would try to feed the inventory with more landslides related to other rainfall events and use them for validation. Also, other roads that were affected by landslides for the same rainfall event could be appropriate to validate the model. And I acknowledge that the collection of this type of data is not straightforward, so if it’s not possible, another option could be to split the road stretch in several portions and apply a spatial validation approach.
A: We like the idea of cross-validating the model by splitting the road into several, non-overlapping segments for validation and training. We will describe and document this approach in a revised version of the manuscript.
_________________
R: To finish I’d like to recognise the big work that has been done to conduct the research presented in this manuscript and I find the proposed overall methodology very interesting, which probably can provide with promising results soon. Therefore, I encourage the authors to consider the comments in this revision so they can help to improve the work done. In addition, I briefly list some other minor suggestions.
A: We highly appreciate the referee's comments and will carefully address them (as well as the comments below) when revising our manuscript.
_________________
- State very clearly the objectives of the research in the introduction section
- Provide the dates of the surveyed Google Earth images
- Provide the exact period that the cumulated rainfall maps represent
- When you talk about specific tools, such as TopoToolbox, specify in which package it works
- If you resample raster maps, specify the method used
- Add references about Akaike Information Criterion, the Receiver-Operating Characteristics or the Area under the Curve
- In Fig. 1 add Uttarakhand region in the location box
- In Fig. 3 add a scale to the maps
- I would also show the map of the resulting model so the reader can see, visually, the matching of the modelled susceptibility of the road and the landslide occurrences
Citation: https://doi.org/10.5194/nhess-2022-295-AC2
-
AC2: 'Reply on RC2', Jürgen Mey, 04 Apr 2023
Status: closed
-
RC1: 'Comment on nhess-2022-295', Anonymous Referee #1, 08 Feb 2023
This manuscript takes the Rishikesh Joshimath (NH-7) highway in North Akandu, India, as the research area or line. Based on field survey, after the rainfall in September and October 2022, the location information of more than 300 landslides was obtained along the road line. It is helpful to understand the mechanism analysis and spatial distribution law analysis of rainfall landslides in this area, and landslide disaster prevention and mitigation. And this manuscript has high efficiency and good timeliness. However, there are some major defects in this manuscript. The specific suggestions are as follows: (1) The means of data acquisition in this manuscript are very limited, and it is far from enough to rely only on field survey. This can only be limited to some targets, and cannot reflect the overall landslide disaster situation. It is suggested that the author consider introducing multi-temporal satellite images. Moreover, relying only on the field survey, it can not be objectively judged that there is an inevitable relationship between these landslides and the rainfall in 2022. (2) The results obtained from incomplete data are often inconsistent with the actual situation. The statistical analysis results in this manuscript are affected by the limitations of landslide investigation technology. It is undoubtedly very difficult to reach the conclusion that "slope, rainfall amounts, and lithology are the main environmental controls on landslide occurrence", and lack of reliable model support. (3) Because this model is too local regional and so limitation of landslide data, it is difficult to popularize and apply it in other regions, and it is also difficult to obtain more general results. To sum up, due to the inadequacy of the underlying data and methods, I suggest reject the manuscript.
Citation: https://doi.org/10.5194/nhess-2022-295-RC1 -
AC1: 'Reply on RC1', Jürgen Mey, 08 Feb 2023
We thank the reviewer for the comments on our manuscript. We summarize the following points raised by the reviewer (R) and provide our replies (A):
R1: It is far from enough to rely only on field survey for data acqusition as it cannot reflect the overall disaster situation. The authors should consider using multi-temporal satellite images.
A1: We thank the reviewer for the suggestion to use multi-temporal imagery. In fact, we checked each landslide location using imagery made available by Google Earth using the latest and historic imagery (L. 112). We classified each location as new landslide, visible before the Sep-Oct 2022 rainfall anomaly and reactivated landslide. We note that detecting landslides from imagery was not always straightforward since landslide scars are difficult to distinguish from unvegetated engineered slopes and road widening (L. 115).
In fact, we emphasize that our mapping approach enabled us to detect the often small landslides lining the road. Moreover, we are confident that our landslide inventory of (partially) road blocking landslides is near to complete and thus reflects the overall disaster situation quite well.
In a revised version of the manuscript, we will document the visual interpretation of the imagery and its results more carefully.
R2: It can not be objectively judged that there is an inevitable relationship between these landslides and the rainfall in 2022.
A2: We raised this point in the text (L. 185):
"Yet not all field-mapped landslide occurrences can be attributed to the anomalously high rainfall period during September and October 2022. Visually inspecting the locations using Google Earth reveals that 21.4 % of the recorded landslides with road blockages existed before (Figure 1). 17.8 % of the landslides were most likely reactivated by the excessive rainfall because they could not be identified to be road-blocking before the rainfall period. Most landslides (60.8 %) were not identifiable as such in the Google Earth imagery available for several dates before and including March 2022."
We ran the analysis with a subset of 60.8% and the results are very similar to those obtained when taking all landslides into account. In a revised version, we will document this analysis.
R3. Incomplete data leads to results inconsistent with the actual situation. The statistical results are affected by limitations of the landslide investigation technology, and the conclusion that "slope, rainfall amounts, and lithology are the main environmental controls on landslide occurrence" are not substantiated by the model.
A3: We disagree with the reviewer. It is the aim of statistical modelling to infer models/test hypotheses based on incomplete data. We regard the actual pattern of landslides as a point pattern and infer the environmental controls using established statistical models. Our inference is complemented by a careful Bayesian uncertainty analysis (Fig. 4, 6 and 7). Based on our model, slope, rainfall amounts and lithology explain the observed landslide patterns very well. Of course, statistical models enable us to infer correlation, not causation. The mechanistic interpretation of the correlations are subject of the discussion.
R4: The model does not provide more general results and cannot be applied to other regions.
A4: This comment addresses nearly any field-based study. Whether our model can or cannot be applied to other regions is elusive and remains to be tested. The consistency of the individual effects of lithozones with previous assessments of landslide susceptibilities of different Himalayan lithologies suggests, however, that our model is applicable beyond the studied extent of the NH-7.
With kind regards,
Jürgen Mey on behalf of all coauthors.
Citation: https://doi.org/10.5194/nhess-2022-295-AC1
-
AC1: 'Reply on RC1', Jürgen Mey, 08 Feb 2023
-
RC2: 'Comment on nhess-2022-295', Anonymous Referee #2, 30 Mar 2023
The manuscript deals with an interesting and largely discussed (far from being solved though) topic in the landslide susceptibility scientific community, i.e. the inter-relationship between linear communication infrastructures and landslides. Specifically, this work aims to carry out a landslide susceptibility analysis for a ~250 km long stretch of the National Highway 7 (NH-7), India.
The relevance of the addressed topic is well justified and supported by adequate references. The authors propose an innovative perspective by considering the landslides found along the road as network-attached spatial point pattern inspired by the work of Baddeley et al. (2021), which, from my point of view, is a potential line of research.
However, I believe that the manuscript shows some important issues that make it not suitable for publication in its current version. The objectives of the research are unclear, the applied methods are not explained with enough details and the structure of the information provided should be improved.
In the following lines I’ll try to explain the main weak points of the manuscript and how they could be improved in my opinion:
(1) The authors dedicated a big effort conducting a landslide inventory that collects more than 300 failures. I acknowledge the big value of such a database as a proxy of the exposure of the road against landslides. And in this sens, I do believe that the road and its exposure to small landslides could be modeled considering the morphological and environmental features throughout its margins. But, the authors focus the analysis to only road side occurring landslides stating that they "only mapped landslides with runouts affecting the road" (Line 109). So, the dataset is restricted to those observations that strictly has affected the road, limiting the representativeness of the inventory. In my opinion, with this dataset it is not possible to model the occurrence of landslides correctly, since it doesn’t consider the full picture of the landslide occurrences. I would suggest to change the perspective of the research and reformulate it as a road exposure research rather than landslide susceptibility study.
(2) The consideration of dynamic environmental maps, such as the cumulated rainfall, as one of the possible controlling factors is conceptually correct and its effect in the landslide occurrence is justified. However, these kind of inputs should be used with extreme care, and the suitability of rainfall products for spatial modelling have to go always very well supported. As an example, a cumulated rainfall map of 100 km of resolution is not going to be able to represent the rainfall variability within a study area of 250 km². In this work the rainfall distribution of that raining event has been represented by several cumulated precipitation maps, which the authors acknowledge to be considerable different among them. The authors select CHIRPS v2 based on the better performance of the model when this map is used as input, and there are no evidences provided that this map really represents the actual rainfall pattern in this area. In my opinion, the statistical results support the idea that the landslides in the inventory are controlled by the rainfall pattern displayed by CHIRPS v2; but, what we are not certain at all is wether this pattern is what has really happened or not, and if the original resolution of this map represents fairly the rainfall variability at the scale of the studied area. I suggest to include a complete section that deals with this issue, where the authors can justify very clearly why they use each product. This should include the assessment of the distortion of the original product, at its original resolution, with respect to the resampled one.
Another crucial aspect concerning the rainfall variables is the resampling method applied to convert the original rainfall products (in 0.05ºx0.05º to 0.25ºx0.25º of resolution) to the working environment of 30x30 meters of resolution. There is no information about that in the manuscript.
(3) I agree with the authors about the relevance of the slope gradient as a key controlling factor. But looking to the pictures provided in Fig. 2, it seems that the types of landslides in the inventory were triggered in the road cuts, just at a few meters of distance from the road. So, maybe the spatial resolution of 30x30 is not accurate enough to represent in a fair way the slope gradient at such a short distance. In addition, the followed approach to obtain and assign the slope value to the road pixel is not clear (Lines 144-146): Why is the buffer around the road set up to 210 m? What do you mean with the “nearest pixel”?
I suggest to grow up to a finer resolution DEM. CARTOSAT-1-derived DEM with 10x10 m of resolutions seems to be a good option (https://link.springer.com/article/10.1007/s40808-019-00694-9).
(4) The model validation approach should be described with more details. Are trained models validated with independent dataset? If yes, which criterion has been used to split the training and validation?
Since the collected data are exclusively related to one raining event, i.e., the monsoon season of 2022, I would try to feed the inventory with more landslides related to other rainfall events and use them for validation. Also, other roads that were affected by landslides for the same rainfall event could be appropriate to validate the model. And I acknowledge that the collection of this type of data is not straightforward, so if it’s not possible, another option could be to split the road stretch in several portions and apply a spatial validation approach.
To finish I’d like to recognise the big work that has been done to conduct the research presented in this manuscript and I find the proposed overall methodology very interesting, which probably can provide with promising results soon. Therefore, I encourage the authors to consider the comments in this revision so they can help to improve the work done. In addition, I briefly list some other minor suggestions.
- State very clearly the objectives of the research in the introduction section
- Provide the dates of the surveyed Google Earth images
- Provide the exact period that the cumulated rainfall maps represent
- When you talk about specific tools, such as TopoToolbox, specify in which package it works
- If you resample raster maps, specify the method used
- Add references about Akaike Information Criterion, the Receiver-Operating Characteristics or the Area under the Curve
- In Fig. 1 add Uttarakhand region in the location box
- In Fig. 3 add a scale to the maps
- I would also show the map of the resulting model so the reader can see, visually, the matching of the modelled susceptibility of the road and the landslide occurrences
Citation: https://doi.org/10.5194/nhess-2022-295-RC2 -
AC2: 'Reply on RC2', Jürgen Mey, 04 Apr 2023
A: We thank the reviewer (R) for helpful and constructive comments. Below, we provide replies to the comments and detail how we will address them in a revised version of the manuscript.
_________________
R: The manuscript deals with an interesting and largely discussed (far from being solved though) topic in the landslide susceptibility scientific community, i.e. the inter-relationship between linear communication infrastructures and landslides. Specifically, this work aims to carry out a landslide susceptibility analysis for a ~250 km long stretch of the National Highway 7 (NH-7), India.
The relevance of the addressed topic is well justified and supported by adequate references. The authors propose an innovative perspective by considering the landslides found along the road as network-attached spatial point pattern inspired by the work of Baddeley et al. (2021), which, from my point of view, is a potential line of research.
However, I believe that the manuscript shows some important issues that make it not suitable for publication in its current version. The objectives of the research are unclear, the applied methods are not explained with enough details and the structure of the information provided should be improved.
In the following lines I’ll try to explain the main weak points of the manuscript and how they could be improved in my opinion:
(1) The authors dedicated a big effort conducting a landslide inventory that collects more than 300 failures. I acknowledge the big value of such a database as a proxy of the exposure of the road against landslides. And in this sens, I do believe that the road and its exposure to small landslides could be modeled considering the morphological and environmental features throughout its margins. But, the authors focus the analysis to only road side occurring landslides stating that they "only mapped landslides with runouts affecting the road" (Line 109). So, the dataset is restricted to those observations that strictly has affected the road, limiting the representativeness of the inventory. In my opinion, with this dataset it is not possible to model the occurrence of landslides correctly, since it doesn’t consider the full picture of the landslide occurrences. I would suggest to change the perspective of the research and reformulate it as a road exposure research rather than landslide susceptibility study.
A: We agree with the referee that limiting the inventory to road blocking landslides may abate its representativeness for roadside landslides, in general. We emphasize, however, that we used this mapping criterion to cope with the overwhelming amount of landslides and to be able to take into account the landslides that detached most recently. Also, construction works for road-widening often stripped off vegetation so that differentiating between actual, non-blocking landslides and excavated slopes was not straightforward. In a revised version of the manuscript, we will provide more details on the motivation behind including road-blocking landslides only. Also, we like the suggestion to focus on road exposure rather than landslide susceptibility and reframe the study accordingly.
_________________
R: (2) The consideration of dynamic environmental maps, such as the cumulated rainfall, as one of the possible controlling factors is conceptually correct and its effect in the landslide occurrence is justified. However, these kind of inputs should be used with extreme care, and the suitability of rainfall products for spatial modelling have to go always very well supported. As an example, a cumulated rainfall map of 100 km of resolution is not going to be able to represent the rainfall variability within a study area of 250 km². In this work the rainfall distribution of that raining event has been represented by several cumulated precipitation maps, which the authors acknowledge to be considerable different among them. The authors select CHIRPS v2 based on the better performance of the model when this map is used as input, and there are no evidences provided that this map really represents the actual rainfall pattern in this area. In my opinion, the statistical results support the idea that the landslides in the inventory are controlled by the rainfall pattern displayed by CHIRPS v2; but, what we are not certain at all is wether this pattern is what has really happened or not, and if the original resolution of this map represents fairly the rainfall variability at the scale of the studied area. I suggest to include a complete section that deals with this issue, where the authors can justify very clearly why they use each product. This should include the assessment of the distortion of the original product, at its original resolution, with respect to the resampled one.
Another crucial aspect concerning the rainfall variables is the resampling method applied to convert the original rainfall products (in 0.05ºx0.05º to 0.25ºx0.25º of resolution) to the working environment of 30x30 meters of resolution. There is no information about that in the manuscript.
A: We agree with the referee that spatial rainfall variability is one of the weak points in our study. Clearly, we are limited by a lack of independent validation data which are either unavailable due to restricted accessibility or because they are already included in the development of the rainfall product. The strong difference between the rainfall products also highlights the apparent difficulties in predicting/hindcasting rainfall distributions in these steep mountain areas. As the referee notes, one could use the landslide data to infer the best rainfall product. However, this approach comes with the problem of circular argumentation. Gladly, and as pointed out by us in the manusript, the choice of rainfall distribution affects the overall predictability of road-blocking slope failures, but it doesn't affect the other parameters (slope and lithology). In a revised version of the manuscript, we plan to document better how we dealt with the rainfall data and provide more details on the bilinear resampling technique used to downsample the grids.
_________________
R: (3) I agree with the authors about the relevance of the slope gradient as a key controlling factor. But looking to the pictures provided in Fig. 2, it seems that the types of landslides in the inventory were triggered in the road cuts, just at a few meters of distance from the road. So, maybe the spatial resolution of 30x30 is not accurate enough to represent in a fair way the slope gradient at such a short distance. In addition, the followed approach to obtain and assign the slope value to the road pixel is not clear (Lines 144-146): Why is the buffer around the road set up to 210 m? What do you mean with the “nearest pixel”?
I suggest to grow up to a finer resolution DEM. CARTOSAT-1-derived DEM with 10x10 m of resolutions seems to be a good option (https://link.springer.com/article/10.1007/s40808-019-00694-9).
A: Yes, indeed the spatial resolution of the DEM may be too coarse to study phenomena occurring at finer spatial scales. Unfortunately, photogrammetric DEMs based on CARTOSAT imagery are available only at 30 m resolution and their quality is comparable to other globally available DEMs. We chose 210 m (7 x spatial resolution of the COP30 DEM) to have sufficient pixels to obtain a robust estimates of the slope which can fluctuate strongly in steep terrain. Our approach pixellates the road, and the model requires that the slope values are mapped to the road pixels. Thus, we identified the pixels closest and upslope to the road, and then aggregated slope values measured in up to 7 pixel distance to the nearest road pixels. In a revised version of the manuscript, we will provide more details about this approach.
_________________
R: (4) The model validation approach should be described with more details. Are trained models validated with independent dataset? If yes, which criterion has been used to split the training and validation?
Since the collected data are exclusively related to one raining event, i.e., the monsoon season of 2022, I would try to feed the inventory with more landslides related to other rainfall events and use them for validation. Also, other roads that were affected by landslides for the same rainfall event could be appropriate to validate the model. And I acknowledge that the collection of this type of data is not straightforward, so if it’s not possible, another option could be to split the road stretch in several portions and apply a spatial validation approach.
A: We like the idea of cross-validating the model by splitting the road into several, non-overlapping segments for validation and training. We will describe and document this approach in a revised version of the manuscript.
_________________
R: To finish I’d like to recognise the big work that has been done to conduct the research presented in this manuscript and I find the proposed overall methodology very interesting, which probably can provide with promising results soon. Therefore, I encourage the authors to consider the comments in this revision so they can help to improve the work done. In addition, I briefly list some other minor suggestions.
A: We highly appreciate the referee's comments and will carefully address them (as well as the comments below) when revising our manuscript.
_________________
- State very clearly the objectives of the research in the introduction section
- Provide the dates of the surveyed Google Earth images
- Provide the exact period that the cumulated rainfall maps represent
- When you talk about specific tools, such as TopoToolbox, specify in which package it works
- If you resample raster maps, specify the method used
- Add references about Akaike Information Criterion, the Receiver-Operating Characteristics or the Area under the Curve
- In Fig. 1 add Uttarakhand region in the location box
- In Fig. 3 add a scale to the maps
- I would also show the map of the resulting model so the reader can see, visually, the matching of the modelled susceptibility of the road and the landslide occurrences
Citation: https://doi.org/10.5194/nhess-2022-295-AC2
-
AC2: 'Reply on RC2', Jürgen Mey, 04 Apr 2023
Jürgen Mey et al.
Jürgen Mey et al.
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 627 | 328 | 17 | 972 | 38 | 5 | 8 |
- HTML: 627
- PDF: 328
- XML: 17
- Total: 972
- Supplement: 38
- BibTeX: 5
- EndNote: 8
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
