All Earth

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/tgda21

Proposing an easy-to-use tool for estimating

landslide dimensions using a data-driven
approach

Minu Treesa Abraham, Neelima Satyam, Biswajeet Pradhan & Samuele

Segoni

To cite this article: Minu Treesa Abraham, Neelima Satyam, Biswajeet Pradhan & Samuele
Segoni (2022) Proposing an easy-to-use tool for estimating landslide dimensions using a data-
driven approach, All Earth, 34:1, 243-258, DOI: 10.1080/27669645.2022.2127549

To link to this article: https://doi.org/10.1080/27669645.2022.2127549

© 2022 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group.

Published online: 27 Sep 2022.

Submit your article to this journal

Article views: 1152

View related articles

View Crossmark data

Citing articles: 4 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tgda21
ALL EARTH
2023, VOL. 34, NO. 1, 243–258
https://doi.org/10.1080/27669645.2022.2127549

Proposing an easy-to-use tool for estimating landslide dimensions using a

data-driven approach
Minu Treesa Abrahama, Neelima Satyama, Biswajeet Pradhanb,c,d and Samuele Segonie
a
Department of Civil Engineering, Indian Institute of Technology Indore, Indore, India; bCentre for Advanced Modeling and Geospatial
Information Systems (CAMGIS), School of Civil and Environmental Engineering, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, Australia; cCenter of Excellence for Climate Change Research, King Abdulaziz University, Jeddah,
Saudi Arabia; dEarth Observation Centre, Institute of Climate Change, Universiti Kebangsaan Malaysia, Bangi, Malaysia; eDepartment of
Earth Sciences, University of Florence, Florence, Italy

ABSTRACT ARTICLE HISTORY

The increase in population and urbanisation of hilly regions have increased the risk due to Received 26 May 2022
landslides. This manuscript presents a data-driven approach with a random forest algorithm to Accepted 20 September 2022
estimate the projected area, length, travel distance, and width of landslides, using elevation and KEYWORDS
slope information. The method is tested for two different study areas (Idukki and Wayanad), landslides; hazard; random
using three different combinations of inputs. The input features considered were elevation (E), forest; travel distance;
tangential slope (θ), drop height (H), angle of reach (α) and the profile curvature (c). A total of 144 machine learning
models were considered and were evaluated using mean-absolute-error (MAE) and root-mean-
square-error (RMSE) values. The results indicate that, by using E and θ alone, the RMSE value in
estimating the length values for flow-like landslides in Wayanad was reduced from 472.74 m to
204.64 m. Out of the 48 combinations considered, MAE values have increased in seven cases and
RMSE values in eight cases only. The pre-trained models are saved and used to develop an easy-to
-use tool, which can bypass the complications associated with the existing statistical approaches.
The tool can be used by untrained personnel for preliminary hazard assessment.

1. Introduction
term failure is critical, as it decides the separation of
Landslides are common natural hazards in hilly phases in the process. Failure indicates the first forma­
regions, responsible for severe economic loss and tion of a rupture surface as displacement (Leroueil
casualties across the globe. With the change in climate et al., 1996), and it happens when the factor of safety
and the increased number of extreme rainfall events, (FS) becomes lesser than 1. This stage involves
the number of rainfall-induced landslides has also aProposing an easy-to-use tool for estimating landslide
increased (Gariano & Guzzetti, 2016). The local people dimensions change in kinematic behaviour, from slid­
are the first ones to identify the tension cracks or minor ing to fall or flow. This change is also critical in deciding
failures before the occurrence of a landslide. If the area the post-failure behaviour. This analysis can be carried
that may get affected by the failure can be effectively out either in a forensic style, as a back analysis, or as
communicated to them based on the available infor­ a prediction for future events.
mation, it can be helpful in making the action plan and The displacement post-failure is highly dependent
communicating the same with the stakeholders. In on the type of failure. The knowledge of landslide
a recent study, it was stated that people are more likely typology is critical in analysing the post-failure motion,
to trust early warnings than structural mitigation mea­ and for precise runout analysis, complex process-
sures (Huang et al., 2021). Understanding the failure based models should be used separately for each land­
mechanism and the post-failure movement of land­ slide type (Armento et al., 2008; Guzzetti et al., 2002).
slides is an essential part of the assessment of Many empirical, analytical, and dynamic models have
a landslide hazard. The attempts to understand the been used to quantify the post-failure motion of land­
mechanism of failure and the post-failure motion of slides. The advancements in numerical models have
slope instabilities had started in the early 20th century helped in understanding the triggering and runout
itself (Terzaghi, 1950). The system of landslides and mechanisms of landslides (Christen et al., 2010), yet
their evolutions has different temporal scales, and the the time taken for analysis and the complexities
variations with respect to time involve different stages involved in modelling still makes this a challenging
such as deformations before the failure, the failure task (McDougall, 2017). The complexities associated
itself, and the displacements after the failure. The with runout modelling are primarily due to the lack

CONTACT Minu Treesa Abraham [email protected]; [email protected] Department of Civil Engineering, Indian Institute of
Technology Indore, Madhya Pradesh, India
© 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (http://creativecommons.org/licenses/by-nc/4.0/), which
permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
244 M. T. ABRAHAM ET AL.

of guidance in this regard. The selection of the runout Three different landslide typologies are considered
model and clear guidance on modelling for practi­ (namely flows, slides and falls) and the data from two
tioners are provided in very few codes or guidelines test sites are used for training and testing the model.
(Lato et al., 2016). The runout analysis is still considered The model uses a Random Forest (RF) algorithm, and
a speciality service that demands expert support the trained model is used to develop a user-friendly tool
(APEGBC, 2012). Considering these facts, the practi­ that can be easily used for applications in landslide
tioners need much simpler approaches for hazard hazard assessment. The method is tested for two differ­
assessment. ent study areas in the Western Ghats of India.
Even though there are empirical and statistical corre­
lations used to estimate the landslide deposition area
2. Study area
and runout distances, the relationship demands prior
information on landslide volume in most cases The proposed methodology is tested at two different
(McDougall, 2017). Two widely accepted relationships locations in the Western Ghats of India. The Western
in this regard are the inverse relationship between the Ghats is a mountain range running through the
reach angle and landslide volume and the one between Western coast of India, with a stretch of 1,600 km.
volume and area of landslides, using Galileo scaling The mountain range highly influences the monsoon
laws. Zhao et al. (2022) studied the empirical equations weather patterns in India and is one of the hotspots of
for estimating the runout distances of landslides and biodiversity in the world (Myers et al., 2000). The
have argued that such may not be useful for regions Western Ghats is separated into two parts by
other than the one from which data is collected. They a mountain pass called the Palghat gap (Figure 1). In
have proposed a Bayesian method for estimating run­ this study, two regions in the Western Ghats, one on
out distance from sparse data, using drop height and the northern side and one on the southern side of the
the slope angle. In another study, a data-driven frame­ Palghat gap, are considered for the analysis. Both dis­
work has been developed to predict the runout distance tricts belong to the state of Kerala. The boundaries of
of landslides, using slide width, slide length, slide the study areas are determined by the administrative
volume, slide thickness, and vertical drop (Xu et al., division (district), but in the case of both the study
2019). They have compared five different algorithms areas, the administrative boundaries coincide with
and have found that multi-layer perceptron performs the geographical boundaries defined by the hills and
better than the other algorithms considered. (Mergili valleys of Western Ghats as well. In August 2018, extre­
et al., 2019) combined the release and runout in land­ mely heavy rains triggered landslides and floods in
slide susceptibility modelling using probability density Kerala, leading to a recovery need of 4.4 billion US
functions and cumulative distribution functions of the Dollars (United Nations Development Programme,
travel distances and angles of reach of the historical 2018). Idukki and Wayanad were the worst hit due to
landslides. The method provides approximate the sus­ landslides, and the backward socio-economic condi­
ceptibility of any point in a landscape, to be affected by tions of these districts also put them in a highly vulner­
either shallow landslide processes or the resulting fail­ able condition. Owing to the higher number and
ure triggered debris flows, either through release, or catastrophic effects of landslides in these regions,
through runout (Lima et al., 2019). The exportability of quantitative hazard assessment and identification of
this model to any other study area is subject to a long elements exposed to risk is the need of the hour.
process of precise data collection and statistical analysis, Idukki and Wayanad are major tourist spots in the
and it requires the thorough knowledge of multiple state, and tourism and agriculture are the major income
probabilities, to arrive at an integrated susceptibility source of the inhabitants. Idukki has a relatively flatter
index. Similar to this study, Melo et al. (2019) have also area in the western part of the district, and the remain­
combined both failure and runout of shallow landslides ing locations are covered by highlands. Wayanad, on the
using logistic regression and a cellular automate model. other hand, has hills along the district boundaries. An
This study is an attempt to bypass the intricacies east-flowing river Kabani and its tributaries have con­
associated with numerical modelling and existing corre­ tributed to the landscape development of Wayanad.
lations using a data-driven approach. As an initial step, The tributaries of the river originate in the hilly regions
a simple tool is introduced, that requires only topogra­ on the western side and flow downhill to the lower
phical features derived from satellite-based information elevation parts on the northeastern side of the district.
to estimate the maximum runout length, maximum The lithology of both the regions is made up of
width, maximum travel distance, and area affected rocks of the migmatite group, and the peninsular
due to a landslide, when the source area is identified. gneissic complex. Both Idukki and Wayanad contribute
The method is an integration of geomorphological and to the major forest cover of Kerala, with a forest area of
geometrical landslide runout assessment techniques. 3151 km2 and 1580 km2, respectively (Forest Survey of
The variables to be estimated were quantified for his­ India, 2019). Owing to the thick forest cover, both
torical events using geomorphological assessment. regions are rich in forest soil of high organic content.
 ALL EARTH 245

Figure 1. Location of study areas. a) India, b) Wayanad, and c) Idukki.

The midlands of these districts have lateritic soil cover, 2021b) for both the study areas, and the procedure is
formed by the transportation of weathered rock. mentioned in the ‘data collection’ section. The land­
Riverbanks have alluvial deposits. While Wayanad and slides were categorised into shallow landslides, flows,
the northern parts of the Palghat gap are characterised and rockfalls to train the models separately (Varnes,
by thick regolith deposits, Idukki has a lesser thickness 1978). The topographical details were collected using
of overburden soil. The high elevation ranges of both the digital elevation models (DEMs) of the study areas
the districts are highly dissected and have witnessed (Alos Palsar DEM, with 12.5 m resolution (ASF DAAC,
deep-seated movements in history. Major debris flows 2015)), and the maximum length, maximum width,
extending up to a few kilometres have happened in area affected, and maximum horizontal distance was
both districts. In Idukki, the road networks in the dis­ measured for each landslide polygon to get the train­
trict where the unsupported vertical slopes are highly ing and testing data. The prepared data were used for
affected by landslides. In the case of Wayanad, most training a model using the RF algorithm and the
landslides have happened within the forest areas. trained model was then used to develop a user-
From the interaction with the local people of the friendly tool for estimating landslide dimensions. The
study areas, it was understood that tension cracks or steps involved in the methodology are explained in
new streams are usually observed prior to failure in the detail in the following sections.
landslide location, particularly in the case of failure-
triggered debris flows. This study is an attempt to esti­
mate the landslide dimensions based on the topogra­ 3.1. Data collection
phical features of the source area using a simple and
Preparation of landslide inventories is the principal
easy-to-use tool. Once the tension cracks are identified,
data required for the development of the model. For
the proposed tool can estimate the runout area, thus
both Idukki and Wayanad, the major landslide disaster
helping to identify the elements exposed to risk and
that happened in 2018 was considered for the analysis.
taking necessary precautions and emergency measures.
Landslides happened throughout the higher elevation
regions of both the districts in 2018, and satellite
images before and after the disaster were available
3. Methodology
for preparing the inventory. The fast vegetation
The study proposes a user-friendly tool for estimating regrowth in the region might have resulted in missing
landslide dimensions based on a data-driven some of the events, yet the prepared inventory was
approach. A detailed landslide inventory was prepared found to be in good agreement with the point land­
using Google earth images (Abraham et al., 2021a, slide inventory prepared using field investigations and
246 M. T. ABRAHAM ET AL.

satellite data interpretation (Hao et al., 2020). The pro­ the values of FS fall below 1. The shape of the slip
cess of preparing the inventory is shown in Figure 2. surface, the material involved, and the topographical
A total of 2162 landslides from Idukki and 388 land­ conditions decide the post-failure motion. From the
slides from Wayanad were mapped using the approach collected data, no cases of topples and spreads were
mentioned in Figure 2. After locating and mapping, the detected, and the inventory was classified into falls,
typology of landslides is evaluated in detail by inter­ slides, and flows. The flows are characterised by long
preting the google earth images. Based on the type of runout, often channelised and flowing towards
failure, the landslides can be classified into five a stream downstream. The flow-like landslides are
(Figure 3a). The initial classification by Baltzer in 1875 usually composed of both soil and rock, and the mate­
(A, 1875) had only three categories: fall, flow, and slide, rial can be classified as debris. Such flows are failure
and this was later modified with the addition of topple triggered, with a translational or rotational slide at the
and spread. crown area and then progressing as a flow due to very
As explained in Figure 3b, the slope fails when the high moisture content. Even though they are complex
driving forces exceed the resisting force, that is, when failures, including both slide and flow, the term ‘flow’ is

Figure 2. Preparation of landslide inventory data from pre and post landslide Google Earth Images. a) Image before the landslide,
b) Image after the landslide, and c) Elevation profile along the landslide body.

Figure 3. Types of failure and post-failure movement. a) types of failure (modified after ; Cruden & Varnes, 1996))and b) Illustration
of landslide process with respect to time (modified after ; Terzaghi, 1950)).
 ALL EARTH 247

used to classify such landslides, indicating the post- Figure 4. The term L denotes the projected runout
failure motion. Slides were identified where earth or length measured through the centre of the landslide
debris are exposed, with lesser runout and clear dis­ body in the plan, and the term W denotes the max­
tinction of a failure plane. The failure of rock in the imum width of the cross-section of the failure. Even
study areas is often characterised by a complex form of though W is also used in this manuscript to represent
both sliding and fall. The disintegrated particles of rock the region Wayanad, the name of the dataset will
fall and travel longer distances and hence the com­ always be used along with a second letter or number,
bined slides and falls of rock are categorised as ‘falls’ in representing the type of failure, and hence both can be
this study. distinguished easily. The projected distance in the plan
A total of 12 datasets were prepared with the col­ between the source and the farthest deposit location is
lected data for training the model. This includes four termed the projected travel distance, and in this study,
sets of data for each study area, one for each landslide it is denoted as D, as shown in Figure 4. Using D and W,
type and one without separating the landslide type. a rectangular bound can be proposed, which can be
Apart from the separate dataset for each study area, considered as the maximum area that can get affected
one common dataset is also prepared for each land­ by the landslide. The drop height ðH) is defined as the
slide type and one superset of all the landslides from projected distance between the source and the farth­
both the study areas. est deposit location on a vertical plane. Other features
These were named I1, I2, I3, IC, W1, W2, W3, WC, C1, in Figure 4 are defined (by ; Hungr et al., 2005) as reach
C2,C3, and CC, where the first letter stands for the angle (α), shadow angle (β), source-talus angle (Ψ), and
region and the second one stands for the type of fail­ substrate angle (γ). The term Dx is the component of D
ure considered. The letter I stands for Idukki, W stands in the global direction parallel to the slope. Apart from
for Wayanad, and C stands for the combined dataset. In these features, the tangential angle made by the slope
the second part, 1 represents flows, 2 represents slides, area to the vertical plane is denoted as θ: For different
3 represents falls and C represents the combined types of slopes, Finlay et al. (1999) have proposed the
dataset. expressions to calculate D as a function of vertical
drop, slope angle, volume and width of landslide. In
the cases except for cut slope and boulder fall, prior
3.2. Terminology and selection of features
knowledge of landslide volume is required to estimate
The main objective of the tool is to minimise the the travel distance.
number of features used for estimating landslide The extent of landslides is highly affected by the
dimensions. As the soil of both the study regions is of geometrical features, but the distances and angles
varying grain size, debris or soil is involved in both about which the information is available before the
slides and flows, while the category falls includes rock landslide can only be considered for predicting the
particles only. Apart from the material, the topographi­ landslide dimensions. The features H, α; θ and c can
cal features play a critical role in deciding the dimen­ be calculated from the DEM data and can be used for
sions of a landslide. The path of flow-like landslides is the estimation of D,W, L; and A. Both H and α require
highly influenced by the topography, ridges, and val­ the knowledge of the nearest flatter area or approx­
leys. From the visual interpretation of Google Earth imate possible runout area. In the case of flow and falls,
images, the travel distances and the area (A) affected the post-failure movement will not stop at the first
by landslides are measured manually, as shown in drop. To summarise, five input features are used, and

Figure 4. Geometrical features associated with landslides: a) plan, and b) longitudinal section (modified after ; Hungr et al., 2005)).
248 M. T. ABRAHAM ET AL.

it is explored in detail if the dimensional parameters 3.3. Machine learning algorithm and
can be predicted effectively using these variables. performance evaluation
The distribution of all four variables in different
RF is a widely used ensemble machine learning (ML)
datasets is summarised in Figure 5. It can be under­
algorithm (Ho, 1995). As the name indicates, a large
stood that the A; L and H values are much higher for
number of decision trees are involved in the decision-
flows in Wayanad, when compared to all other cases. In
making process of RF. Each tree in an RF has multiple
the case of W, slides have higher values than other
branches and nodes. At each node, a decision is taken,
datasets. W has the least varying distribution among all
which leads to one of the branches. The decisions thus
the variables.
For each of the 12 datasets prepared, three different continue, considering all the features, and the tree
trials were conducted, by varying the combination of finally assigns a class to the object. Each tree will
input features, to predict all four variables (A; L; W; and have a separate prediction, and later, the final predic­
D). In the first combination, all five features are con­ tion is decided based on voting, considering the pre­
sidered and is named EHθαc. The feature importance dictions of all decision trees (Figure 6). Each decision
values of each feature are used to understand the tree is sampled independently using statistical boot­
significance of features in the combination. Feature strapping (Breiman et al., 2003) and contains a subset
importance calculates a score for all input features, of the dataset considered.
which represents the significance of each feature. The RF is widely used for multiple applications to train
value of feature importance can vary from 0 to 1, and models and is proven to provide satisfactory results
a higher value indicates that the specific feature has due to the random selection at nodes. The method is
more effect in predicting the variable. As the objective ideal for minimising the overfitting issues. The perfor­
is to minimise the number of features, based on the mance of the model can be further fine-tuned by
feature importance values obtained for the first com­ varying the set of hyperparameters. The performance
bination, two more combinations were considered, of the model is highly sensitive to the values of hyper
one with E; H and θ, and the last one with only two parameters (Daviran et al., 2021), and several cost-
features, E and θ. effective ways are available, for multi-criteria optimi­
Each dataset was trained and tested separately sation (Liu et al., 2017). The number of trees in the
using the RF algorithm to find the best-suited features forest, the maximum number of features considered
for estimating landslide dimensions. for splitting a node, the maximum depth of the tree,

Figure 5. Box and whisker plot showing the distribution of A; L; W and D in different datasets.
 ALL EARTH 249

Figure 6. Graphical representation of RF algorithm.

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and the minimum number of samples used to split 1X n �2
a node are varied in this study to improve the effi­ RMSE ¼ ypred;i ytest;i (3)
n i¼1
ciency of the model. These values were fine-tuned
separately for each model, along with the test-to-
where n is the total number of samples in the test
train ratio of the dataset. The process is carried out
dataset, ytest;i is the variable in the test dataset and
manually, by varying the parametric inputs and
ypred;i is the corresponding value predicted by the
observing the corresponding model performance.
model. Also, ytest indicates the mean value in the test
All the other parameters are constant while fine-
dataset. The value of d varies from 0 to 1, and the
tuning one parameter. In this study, RF regressor is higher the value, the better the agreement between
used to predict the variables A; L; W and D using three observed and predicted values. Based on the d, MAE
different combinations of input features. The test-to- and RMSE values, the features were decided and the
train ratio was fixed for each of these variables, based tool for landslide dimension estimation was developed
on the performance of the EHθαc model. This is to by using the best-suited feature inputs.
ensure that the comparisons of errors are made on
the same dataset. The ratio was varied from 0.1 to 0.5
in the case of the EHθαc model, and the best perform­ 4. Results
ing test dataset was used to test the performances of
the other two cases as well. Different trials were con­ Considering the regional-specific and failure-specific
ducted to understand the effect of each feature, and datasets and different feature combinations, a total of
the Willmott’s index of agreement (d), Mean-Absolute 144 models were considered for the comparative ana­
-Error (MAE) and the Root-Mean-Square-Error (RMSE) lysis. As expected from the geometrical properties,
values of the predicted and observed values in the when more specific input features are available, the
test dataset were used to evaluate the performance of prediction performance will be improved. But the pos­
different models. These values are calculated based sible information available prior to failure is highly
on the test and predicted data, using the following limited to the scarp zone, where tension cracks are
equations: limited. When multiple flat areas are available along
the failure propagation, finding out the values of H and
Pn � �2 θ is challenging. The first step is to understand the
i¼1 ytest;i ypred;i
importance of each of these features in predicting
d¼ 1 Pn ��� ��2 (1)
� the variables. The least important features can be
i¼1 �ypred;i ytest jþj ytest;i ytest �
removed from further analysis. When all the features
have similar importance values, it is crucial to under­
n � stand the sensitivity of each feature by understanding
1X �ypred;i
��
MAE ¼ ytest;i � (2) the variations that may happen in the prediction per­
n i¼1
formance if it is removed.
250 M. T. ABRAHAM ET AL.

4.1. Feature importance and performance consider when looking for the best split were fine-
evaluation tuned, keeping default values for all other parameters.
Considering the lesser number of parameters and
The feature importance values of all 144 combinations trials, the fine-tuning was carried out by varying the
are shown in Figure 7. The values represented in values manually, for each dataset, and each feature
Figure 7 are the feature importance values obtained combination. The value of maximum depth of trees
for the fine-tuned models. The maximum depth of were varied from 5 to 500, with an increment of 50,
trees, number of trees, minimum number of samples and it was observed that the performance of training
used to split a node, and the number of features to data went increasing, while the performance of test

Figure 7. Feature importance values for all features for 12 datasets considered for predicting a) A, using EHθαc, b) A, using EHθ, c)
A, using Eθ, d) L, using EHθαc, e) L, using EHθ, f) L, using Eθ, g) W, using EHθαc, h) W, using EHθ, i) W, using Eθ, j) D, using EHθαc, k)
D, using EHθ, and l) D, using Eθ.
 ALL EARTH 251

data was decreasing beyond the fine-tuned value. For will not make much variations in α value, but this is
the number of trees, the values were varied from 50 to significant in the case of slides. Estimating α prior to
1000, with an increment of 50. The observation was failure is an almost impossible task, yet most of the
similar for both training and testing datasets, and the existing correlations use α as an input parameter for
fine-tuned value is the value beyond which no signifi­ predicting D. From the feature importance values of
cant improvement in model performance is noted. The the combination EHθαc, it was observed that α is
number of samples at each node were varied from 2 to highly significant in only a few cases, and in all the
10, with an increment of 1. For the number of features, other cases, it is neither the least nor the most signifi­
the values were varied between 1 and 3, with an cant factor. Hence, in the second combination, α was
increment of 0.5. All these increments were further also removed along with c.
reduced once the performance of test data becomes From the second combination, it was observed that
constant or starts decreasing. Fine intervals were used E and θ have very high feature importance values in
in such cases to decide the model parameters. All 144 some cases, while H has values comparable to the
models considered in this study have separate fine- other two, even when it is the most significant factor.
tuned parameters, and the feature importance value, Also, H is least important in predicting D in the case of
MAE and RMSE of the fine-tuned models were further I3, C3 and CC. Even though drop height can be esti­
used for comparison. As shown in Figure 7, the curva­ mated from the profile of the slope before failure, and
ture values are least important in all the trials con­ it is challenging when the slope has multiple landings
ducted using EHθαc. While all other features have in between. It is difficult to use the value of H before
feature importance values greater than 0.15, the corre­ failure. A wrong value can lead to large variations in
sponding values for c were found to be less than 0.1 in predicted and observed dimensions of landslides.
most cases. The value has slightly gone above 0.1 only Hence, the third attempt was the combination of
in the case of W3 and C2.This indicates that the feature only E and θ, which can be easily calculated once
c is less significant in predicting the variables, and the tension cracks are observed. The feature importance
results may not get affected highly even if this input is values in this combination indicate that both the fea­
avoided from the analysis. Hence, c is removed from tures have importance values close to 0.5 and are more
further analysis. or less equally contributing to the prediction of vari­
While considering the features, the feature impor­ ables, in most cases except C3 for predicting W, and I1
tance values of E and θ were found to be more than 0.3 for predicting D. In the case of C3 for predicting W; θ
in some cases. The effect of θ was found to be more was found to be more significant with an importance
significant in the case of C1 for predicting A (0.30), L factor of 0.64 and in the second case, E was found to be
(0.30) and D (0.28), C2 for predicting A (0.32), C3 and C4 more significant with an importance factor of 0.62.
for predicting L (0.27 and 0.27) and D (0.28, 0.30). Apart from deciding the combinations, the predic­
Similarly, the effect of E was significant in the case of tion performance of all three combinations should be
IC (0.31) and WC (0.29) for predicting W, and I2 (0.26), evaluated in detail. This helps in understanding the
I3 (0.28) and C2 (0.28) for predicting D. In the case of applicability of the model for the intended purpose.
W3 for predicting A, the feature E had the second least As the EHθαc combination has more controlling para­
importance value of 0.15, where α was found to be the meters, it is expected to provide better prediction
most crucial with a feature importance value of 0.27. performance. If the results of the other two combina­
From the results, it was observed that E and θ have tions are highly varying from the performance of this
significant effects in predicting the landslide dimen­ combination, the model cannot be used effectively
sions, particularly when combined datasets are used. with less number of features. Even though H and α
This has special significance, as the model has to be were not the most significant factors in most cases, and
exported to multiple regions after a detailed testing their importance values were comparable with those
procedure, and this is the first step in developing of E and θ. Hence, removing these parameters can
a globally applicable model. Hence, when the model critically affect the predicted values. This has been
is applied to a new region for which trained models are evaluated using MAE and RMSE values, as shown in
not available. It is suggested to use the combined Figure 8. Along with the error values in the first com­
dataset for predicting the dimensions. This has led to bination (EHθαc) represented using the line diagram,
the inference that E and θ cannot be avoided while the percentage variation of the error values obtained
developing a model. The variables H and α decide the by using the second (EHθ) and third (Eθ) are repre­
value of D. If both H and α are known, D can be easily sented as a scatter plot in Figure 8.
calculated using the geometry without using any pre­ From Figure 8, it can be observed that the errors
diction model. The challenge in this regard is the diffi­ and significantly large for W1 and WC, where long
culty in understanding the value of α. When the terrain runout debris flow with very large areas are dominat­
has a D value extending up to a few kilometres, minor ing. The increased values of error are in proportion
variations in the expected farthest deposition point with the values in the database and hence are
252 M. T. ABRAHAM ET AL.

Figure 8. MAE and RMSE values in predicting different variables. a) MAE in predicting A, b) RMSE in predicting A, c) MAE in
predicting L, d) RMSE in predicting L, e) MAE in predicting W, f) RMSE in predicting W, g) MAE in predicting D, and h) RMSE in
predicting D.
 ALL EARTH 253

Figure 9. The interface of the proposed tool. a) All inputs, b) dropdown for region, and c) dropdown for landslide type.

acceptable. The primary concern here is the variation known, the corresponding option can be selected,
in the error values with the change in the input feature. and NA can be selected if the type is unknown. When
When the percentage variation is negative, it indicates the material is soil or debris, and the moisture content
that the error is decreased and; therefore, the perfor­ is less, or when the terrain is relatively flat with pre­
mance is improved after removing the features. When dominantly cut slopes, slides can be expected. If the
the variation is positive, it has a negative impact on the terrain has a very high moisture content and spring
prediction. After evaluating all the cases, it can be formations are observed nearby, flows can be
understood that the prediction is affected by the expected, and falls can be expected only in the case
change in features, and the maximum value of percen­ of rocks. Based on the selection, the corresponding
tage variation was observed, −56.92%, in the case of datasets among the 12 will be selected. The selection
estimating L of W1. The reduction has happened while of the dataset has been decided after comparing the
using EHθ combination. The corresponding reduction error patterns, as shown in Figure 10.
in error using Eθ combination is −56.71%. The RMSE As observed from Figure 10, even though Wayanad
values of L in these two cases are 203.65 m and has its own regional-specific database, the combined
204.64 m, respectively. This indicates that the error dataset has lesser values of error in most cases. With
has been reduced considerably when the α and c are the use of the combined dataset, the error has
removed. The maximum increase in error is 11.63%, in increased while predicting width, in all the cases,
the case of MAE, for predicting D for W3 dataset, using with the maximum increase of 70.82% of RMSE value
EHθ. The corresponding variation while using Eθ is in the case of slides. The minimum increase is for MAE,
−10.57%. The results indicate that in no case the error for flow-like landslides, which is 7.55%. Also, in the case
has increased beyond 11.63%, even after removing the of slides, both MAE; and RMSE values have increased
features. This is an acceptable limit, particularly in the with the usage of the combined dataset. Hence, when
case of long-runout events like falls and flows. Based landslide type is selected as slide, the trained model for
on the results, Eθ combinations of all cases were saved Wayanad, W2 is used. The models with the Wayanad
into pickle files as predictors to develop an interactive database are also used for predicting W when the
tool for predicting the landslide dimensions. region is selected as Wayanad. But in all other cases.
The combined dataset is selected for predicting the
dimensions of landslides in Wayanad to minimise the
4.2. Description of the tool
error. In the case of Idukki, the error is less when the
The tool is designed with the objective of delivering an regional-specific database is used. Hence, for Idukki,
easy-to-use platform for the practitioners to estimate the I1; I2; I3 and IC datasets are used, according to the
the area that may get affected by landslides and take selection of user, and for Wayanad, the combined
necessary actions. The tool is completely developed in dataset is selected for flows, falls and when the type
python environment, using the existing library func­ of failure is unknown. When the input for region is NA,
tions. The regression is carried out using scikit-learn the combined dataset is selected, corresponding to the
(Pedregosa et al., 2011), and the trained models are failure type. When both region and landslide type are
saved using pickle (Van Rossum, 2020). The tool has an NA, CC dataset is used for prediction.
interface developed using tkinter (Moore, 2018), which After selecting all the inputs, the ‘Calculate’ button
requires the elevation and tangential slope of the can be used to get the outputs displayed on the
region as input features (Figure 9). screen. The output variables can be used to estimate
The two subsequent inputs decide which model the area that may get affected by the hazard. The tool
should be used for predicting the results. The user is straightforward to be used and delivers the results
can select Idukki and Wayanad in the present version, within fractions of a second, as it uses pre-trained
and if the region is outside these two, the NA option models for prediction. The performance of the tool is
can be selected. Similarly, if the type of failure is currently evaluated for the datasets retrieved in this
254 M. T. ABRAHAM ET AL.

Figure 10. Comparison of MAE and RMSE values of Eθ combination, for different landslide types. a) MAE for flow, b) RMSE for flow,
c) MAE for slide, d) RMSE for slide, e) MAE for fall, f) RMSE for fall, g) MAE for combined landslide types, and h) RMSE for combined
landslide types.
 ALL EARTH 255

Figure 11. MAE, mean and d values of A; L; W and D in different datasets.

study only. For extending its applicability other collecting data from multiple regions in order to
regions, thorough analysis with regional specific data develop a globally applicable version of the tool.
is required. While the existing literature presents sta­
tistical correlations with drop height and other para­
5. Discussion
meters which are obtained after the occurrence of
landslide. Such relationships have limited applicability Study presents a data-driven approach to predict the
on decision support and early warning applications. dimensions of different landslide types using an easy-
This study puts forward a set of promising results to-use tool. The results indicate that the geometrical
which indicate that the dimensions of landslides can parameters such as E; H; θ; α and c can be used for
be predicted using the elevation and slope information estimating the area, maximum length, maximum
of the location of visible cracks. The method is much width and maximum travel distance of landslides.
simpler when compared with the numerical model­ The study investigates in detail the possibilities of
ling-based tools and can be used to identify the ele­ predicting the dimensions using the elevation and
ments exposed to risk and take necessary actions slope data alone. Figure 7 indicates that the use of
before the occurrence of landslide. The primary reason the Eθ combination reduces the error when compared
for considering the tool an ‘easy-to-use’ one is the to the EHθαc combination, in most of the cases. While
usage of minimum parametric inputs, with satisfactory evaluating the feature importance values, it can be
outputs, as demonstrated by the MAE and RMSE understood that the curvature is the least important
values. While using a numerical model requires factor in all the 144 combinations considered. This is
detailed knowledge of the triggering mechanism, because most landslides have concave profile curva­
topography, boundary conditions and material proper­ tures, and the values are highly similar irrespective of
ties, this tool can be used easily with only elevation their size. Hence, any change in curvature values will
and slope information. Also, the interface provides an not affect the model performance. Along with curva­
easy way to input the parameters and directly getting ture, the angle of reach, α, was also removed in
the outputs, rather than using regression equations or the second analysis. Even though there was
codes. The method can be further enhanced by a significant reduction in error in some of the cases
256 M. T. ABRAHAM ET AL.

using the second combination, the MAE value has are observed. Considering d, the values are least in
increased by more than 11% in the case of predicting case of slides, and the minimum value observed is
D with EHθ. Eventhough an influential parameter was 0.38, in the case of slides in the combined dataset. In
removed, the performance was not highly affected, as all cases except slides, the values are greater than 0.5,
the hyperparameters were fine-tuned separately for showing satisfactory agreement between the
each model. The reason for E being a critical factor in observed and predicted datasets.
estimating the dimensions is closely related to the The width of landslides is maximum for both slides
landslides that have happened in the study area. In and falls, and the values range from 8 m to 337 m. The
both the study areas, E values are closely related to H: error values should be evaluated with reference to the
The landslides in higher elevation zones have hap­ distribution of data in each case. From Figure 8 it can
pened in forest areas and plantations, where the H be understood that in the case of area, the maximum
values are also higher. The failures in lower elevation MAE is with W1, which is 6269.79 m2. This error has
regions are induced by the cut slopes exposed with­ highly influenced the combined dataset of Wayanad as
out lateral support. Their dimensions are controlled well, which has an MAE value of 5778.57 m2. Even
by the construction activities such as buildings and though the magnitude is higher when compared to
roads, and they usually have lesser drop heights. the other values, the predictions are satisfactory, as the
Thus, even though H values are removed from the area of debris flows in this region ranges from 485 m2
input features, E values indirectly represent H. The to 253,880 m2. The pattern of variation of error is
reason for using E instead of H is that E value can be similar for A; L; and D, with the maximum error in the
collected when the source area is known, while H case of W1 and WC datasets. But in the case of width,
value cannot be. This limits the applicability of the the maximum error is observed in the case of I2 and IC
model to different regions, where E and H are not datasets. This is due to the higher number of cut slope
closely related. Thus, the methodology has to be failures that happened in the Idukki district. The fail­
tested before exporting to other regions. The number ures are very wide and have a width values upto 337 m,
of trees, features at each node, minimum number of with lesser values of length. Also, there are shallow
samples used to split a node and depth of trees were landslides that happened away from the road, which
used to minimise the error in each case. A similar have width-to-length ratio close to unity. The error is
approach was adopted and was found satisfactory in slightest in the case of rockfalls, where the average
reducing the error for Eθ combination as well. This width is 48 m in the case of a combined fall dataset,
helped in developing a model by using only slope and and the values vary from 12 m to 137 m. These obser­
elevation data for predicting landslide dimensions. vations indicate that when the dataset is uniform,
While comparing the errors, it should be noted that there are higher chances that the error will be mini­
the AL; W; and D values are entirely different for each mum. The hypothesis during formulating the metho­
landslide type. From Figure 11, it can be observed that dology was that when the same type of failure
the MAE values in all cases are lesser than the mean happens in a nearby location, the dimensions will not
values of the dataset. While flows and rockfalls have vary much, and hence a model trained using historical
very long runouts extending from a few hundreds of data can effectively be used for predicting future
metres to a few kilometres, slides extend up to a few events in the same area.
hundreds of metres only, with an average value close In the case of the Wayanad dataset, this hypothesis
to 100 m (Figure 11). The MAE values in the case of the is not valid, as the error is lesser for the combined
area are more than 2000 m2 in all the cases, except falls dataset than the regional-specific Wayanad dataset.
for Idukki and the combined dataset. The MAE values in The main reason for this variation is among the 388
predicting the length of the slides are also very close to landslides mapped from Wayanad, 252 are flows, 68
the mean values of the dataset. This has happened due are slides, including both earth and debris, and the
to the considerable variation in the size of slides. The remaining are falls. The number of events that is
error values in prediction will overestimate the land­ being used for training is much less when compared
slide hazard in the case of small landslides. This limita­ to the whole dataset, and the variation within the
tion can be bypassed by preparing a different dataset dataset is also very high. Due to these reasons, the
for cut slope failures and shallow landslides in other model gets better trained with combined datasets.
locations. While shallow slides happening in forest Hence, the regional-specific datasets need not be the
regions have a wider travel distance, the failed mass best option while predicting landslide dimensions, and
in case of cut slope failures often gets deposited at the the number and quality of data plays critical roles in
foot of the slope due to the flat area nearby. In such minimising the error. With more data, the model can
cases, the failure plane is primarily vertical, while in the be extended to other parts of the world as well.
case of other shallow landslides, circular or transla­ The methodology can only be used to predict the
tional slip surfaces inclined to the horizontal plane dimensions of a landslide that may happen in the
 ALL EARTH 257

future, based on the field observations of cracks. Data availability statement

However, no information can be provided on the
The data used for the analysis is available on request from
time of occurrence of the landslide. The tool should the corresponding author.
be the regional or local scale landslide early warning
systems to obtain the information on ‘when’
a landslide will occur. The warning system can be Tool availability
based on rainfall thresholds, seismic signals, or satellite The tool mentioned in the manuscript will be shared on
or field-based monitoring systems. The integration of request by the corresponding author.
this tool along with an early warning system can pro­
vide a better understanding of the hazard and can be
used to disseminate the warnings effectively to the References
stakeholders. This aspect can be explored in the future. A, B. (1875). Über bergstürze in den Alpen. Verlag der
In the present state, the proposed methodology is best Schabelitz’schen buchhandlung (C. Schmidt). Zurich.
suited for long runout failures like flows and falls and Abraham, M. T., Satyam, N., Jain, P., Pradhan, B., & Alamri, A.
can be used to develop a globally applicable model for (2021a). Effect of spatial resolution and data splitting on
landslide susceptibility mapping using different machine
predicting landslide dimensions.
learning algorithms. Geomatics, Natural Hazards and Risk, 12
(1), 3381–3408. https://doi.org/10.1080/19475705.2021.
2011791
6. Conclusions Abraham, M. T., Satyam, N., Lokesh, R., Pradhan, B., &
Alamri, A. (2021b). Factors affecting landslide susceptibil­
The study presents a data-driven approach to predict ity mapping: assessing the influence of different machine
the dimensions of landslide, upon the identification of learning approaches, sampling strategies and data
minor cracks in the crown area. The methodology splitting. Land, 10(9), 989. https://doi.org/10.3390/
proposed in this study use only the elevation and land10090989
APEGBC. (2012). Professional practice guidelines – Legislated
tangential slope of the crown area to predict the
flood assessments in a changing climate in British Columbia
dimensions of landslides and prove to be a promising (Vancouver, British Columbia: Engineers and Geoscientists
tool that can be used in the decision support system. British Columbia). 2012.
The proposed methodology is tested for two differ­ Armento, M. C., Genevois, R., & Tecca, P. R. (2008).
ent regions in Western Ghats of India, using 12 differ­ Comparison of numerical models of two debris flows in
ent datasets, and three different combinations of input the Cortina d’ Ampezzo area, Dolomites, Italy. Landslides, 5
(1), 143–150. https://doi.org/10.1007/s10346-007-0111-2
features are used to evaluate the influence of each ASF DAAC, 2015. Alaska Satellite Facility Distributed Active
parameter on the model predictions. The comparison Archive Center (ASF DAAC) Dataset: ASF DAAC 2015, ALOS
of MAE and RMSE values of the predicted variables in PALSAR_Radiometric_Terrain_Corrected_high_res;
each case indicates that the maximum increase in error Includes Material © JAXA/METI 2007. [WWW Document].
is only 11.63%, while the reduction in error is 56.92% https://doi.org/10.5067/Z97HFCNKR6VA
Breiman, L., Last, M., & Rice, J. (2003). Random Forests:
with the Eθ combination. The performance of Eθ com­
Finding Quasars Feigelson, E.D., Babu, G. J. Statistical
bination was found to be comparable with the other Challenges in Astronomy (pp. 243–254). Springer-Verlag.
two, without the requirement of any challenging fea­ https://doi.org/10.1007/0-387-21529-8_16
ture inputs like drop height and angle of reach. Christen, M., Kowalski, J., & Bartelt, P. (2010). RAMMS:
The combinations of all 12 datasets were used to Numerical simulation of dense snow avalanches in
decide the model to be used for predicting the dimen­ three-dimensional terrain. Cold Regions Science and
Technology, 63(1–2), 1–14. https://doi.org/10.1016/j.coldre
sions. While the region-specific models were found to gions.2010.04.005
have least errors for Idukki, the combined dataset was Cruden, D., & Varnes, D. (1996). Landslide types and pro­
found to have better performance than the datasets cesses. In A. K. Turner & R. L. Schuster Eds., Landslides,
for Wayanad. The pre-trained models were used to investigation and mitigation (pp. 36–75). Transportation
develop a tool with an interactive interface, which Research Board. Special Report.
Daviran, M., Maghsoudi, A., Ghezelbash, R., & Pradhan, B.
can be easily used to predict the landslide dimensions.
(2021). A new strategy for spatial predictive mapping of
The proposed methodology has the potential to be mineral prospectivity: Automated hyperparameter tuning
applied to other regions as well with the availability of random forest approach. Computers & Geosciences, 148,
of regional specific data, yet the possibility of finding 104688. https://doi.org/10.1016/j.cageo.2021.104688
robust relationships among the variables should be Finlay, P. J., Mostyn, G. R., & Fell, R. (1999). Landslide risk
evaluated through detailed analysis. assessment: Prediction of travel distance. Canadian
Geotechnical Journal, 36(3), 556–562. https://doi.org/10.
1139/t99-012
Forest Survey of India. (2019). State of forest report 2019. In
Disclosure statement Dehradun 2 . Uttarakhand: 131–140 .
Gariano, S. L., & Guzzetti, F. (2016). Landslides in a changing
No potential conflict of interest was reported by the climate. Earth-Science Reviews, 162, 227–252. https://doi.
author(s). org/10.1016/j.earscirev.2016.08.011
258 M. T. ABRAHAM ET AL.

Guzzetti, F., Crosta, G., Detti, R., & Agliardi, F. (2002). STONE: McDougall, S. (2017). Landslide runout analysis — Current
A computer program for the three-dimensional simulation practice and challenges. Canadian Geotechnical Journal,
of rock-falls. Computers & Geosciences, 28(9), 1079–1093. 54, 605–620. https://doi.org/10.1139/cgj-2016-0104
https://doi.org/10.1016/S0098-3004(02)00025-0 Melo, R., Zêzere, J. L., Rocha, J., & Oliveira, S. C. (2019).
Hao, L., van Westen, C., Martha, T. R., Jaiswal, P., & Combining data-driven models to assess susceptibility
McAdoo, B. G. (2020). Constructing a complete landslide of shallow slides failure and run-out. Landslides, 16
inventory dataset for the 2018 monsoon disaster in Kerala, (11), 2259–2276. https://doi.org/10.1007/s10346-019-
India, for land use change analysis. Earth System Science 01235-2
Data, 12(4), 2899–2918. https://doi.org/10.5194/essd-12- Mergili, M., Schwarz, L., & Kociu, A. (2019). Combining release
2899-2020 and runout in statistical landslide susceptibility modeling.
Ho, T. K., 1995. Random decision forests, in: Proceedings of the Landslides, 16(11), 2151–2165. https://doi.org/10.1007/
International Conference on Document Analysis and s10346-019-01222-7
Recognition, ICDAR. pp. 278–282. https://doi.org/10.1109/ Moore, A. D. (2018). Python GUI Programming with Tkinter:
ICDAR.1995.598994 Develop responsive and powerful GUI applications with
Huang, H., Huang, J., Liu, D., & He, Z. (2021). Understanding Tkinter. Packt Publishing.
the public responses to landslide countermeasures in Myers, N., Mittermeier, R. A., Mittermeier, C. G., da
southwest China. International Journal of Disaster Risk Fonseca, G. A. B., & Kent, J. (2000). Biodiversity hotspots
Reduction, 64, 102500. https://doi.org/10.1016/j.ijdrr.2021. for conservation priorities. Nature, 403(6772), 853–858.
102500 https://doi.org/10.1038/35002501
Hungr, O., Corominas, J., & Eberhardt, E. (2005). Estimating Pedregosa, F., Gaël Varoquaux, A. G., Michel, V.,
landslide motion mechanism, travel distance and velocity. Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,
In O. Hungr, R. Fell, R. Couture, & E. Eberhardt (Eds.), Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Landslide risk management (pp. 30). Taylor and Francis Cournapeau, D., Brucher, M., Perrot, M.,
Group, CRC Press. Duchesnay, É., Pedregosa, F., Varoquaux, G.,
Lato, M., Bobrowsky, P., Roberts, N., Bean, S., Powell, S., Gramfort, A., Michel, V., . . . Duchesnay, É. (2011).
Stead, D., McDougall, S., Brideau, M. A., & VanDine, D., Scikit-learn: Machine Learning in Python. Journal of
2016. . In Canadian technical guidelines and best practices Machine Learning Research, 12, 2825–2830.
related to landslides: a national initiative for loss reduction Terzaghi, K. (1950). Mechanism of landslides Paige, Sidney. In
8114 (Geological Survey of Canada)https://doi.org/10. Application of geology to engineering practice. Vol.
4095/299117 . (Geological society of America), pp.83–123https://doi.
Leroueil, S., Locat, J., Vaunat, J., Picarelli, L., Lee, H., & Faure, R. org/10.1130/Berkey.1950.83 doi:
(1996). Geotechnical characterization of slope move­ United Nations Development Programme. (2018). Kerala post
ments. In K. Senneset (Ed.), 7th International Symposium disaster needs assessment floods and landslides-august 2018
on Landslides (CRC Press) (pp. 53–74). (United Nations Development Programme) 1–440.
Lima, P., Steger, S., Netto, A. L. C., Glade, T., & Mergili, M., 2019. Van Rossum, G. (2020 (Python Software Foundation.)https://
Combining landslide susceptibility with potential runout. www.python.org/downloads/release/python-382/). The
An integrative approach combining data-driven methods., python library reference, release 3.8.2.
in: IAG Regional Conference on Geomorphology 2019. 19-21 Varnes, D. (1978). Slope movement types and processes.
September (International Association of Transp. Res. Board Spec. Rep.
Geomorphologists) Athens, p. 536. Xu, Q., Li, H., He, Y., Liu, F., & Peng, D. (2019). Comparison of
Liu, C. H. B., Chamberlain, B. P., Little, D. A., & Cardoso, Â., 2017. data-driven models of loess landslide runout distance
Generalising random forest parameter optimisation to estimation. Bulletin of Engineering Geology and the
include stability and cost, in: Y. Altun, K. Das, Environment, 78(2), 1281–1294. https://doi.org/10.1007/
T. Mielikäinen, D. Malerba, J. Stefanowski, J. Read, s10064-017-1176-3
M. Žitnik, M. Ceci, & S. Džeroski (Eds.), Machine Learning Zhao, T., Lei, J., & Xu, L. (2022). An efficient Bayesian method
and Knowledge Discovery in Databases; Proceedings of for estimating runout distance of region-specific land­
European Conference, ECML PKDD 2017 Skopje, Macedonia, slides using sparse data. Georisk: Assessment and
September 18–22, 2017 Proceedings, Part III, Lecture Notes in Management of Risk for Engineered Systems and
Computer Science. Springer International Publishing, Cham, Geohazards, 16(1), 140–153. https://doi.org/10.1080/
pp. 102–113. https://doi.org/10.1007/978-3-319-71273-4_9 17499518.2021.1952613