Environmental Modelling and Software 170 (2023) 105833

Contents lists available at ScienceDirect

Environmental Modelling and Software

journal homepage: www.elsevier.com/locate/envsoft

Speech-recognition in landslide predictive modelling: A case for a next

generation early warning system
Zhice Fang a, Hakan Tanyas b, Tolga Gorum c, Ashok Dahal b, Yi Wang a, Luigi Lombardo b, *
a
China University of Geosciences, Institute of Geophysics and Geomatics, Wuhan, China
b
University of Twente, Faculty of Geo-Information Science and Earth Observation (ITC), Enschede, Netherlands
c
Eurasia Institute of Earth Sciences, Istanbul Technical University, Istanbul, Turkey

A R T I C L E I N F O A B S T R A C T

Keywords: Traditional landslide early warnings are based on the notion that intensity-duration relations can be approxi­
Landslide prediction mated to single precipitation values cumulated over fixed time windows. Here, we take on a similar task being
Precipitation inspired by modeling architectures typical of speech-recognition tasks. We aim at classifying the Turkish land­
Speech recognition
scape into 5 km grids assigned with dynamic landslide susceptibility estimates. We collected all available na­
Time series
Early warning system
tional information on precipitation-induced landslide occurrences. This information is passed to a Long Short-
Term Memory equipped with the whole rainfall time series, obtained from daily CHIRPS data. We test this
model: 1) by randomizing the presence/absence data to represent the slope instability over Turkey and over 13
years under consideration (2008–2020) and 2) by assessing the effect of different time windows used to pass the
rainfall signal to the neural network. Results show that the inclusion of the full precipitation signal rather than its
scalar approximation leads to a substantial increase in prediction power (approximately 20%). This may
potentially pave the road for a new generation of speech-recognition-based landslide early warning systems.

1. Introduction Osanai et al., 2010). In the context of data-driven models, precipitation

thresholds are mainly estimated by taking a variable selection approach,
Landslide early warning systems (LEWSs, Guzzetti et al., 2020) focus where a classification task is solved using the precipitation measured at
on forecasting slope failures as a function of weather forecast data the date of the landslide occurrence (Melillo et al., 2018). The classifi­
(Gariano and Guzzetti, 2016; Segoni et al., 2018). Regional and global cation performance is then stored, and the second step tests the same
scale LEWSs mainly follow a probabilistic framework (Pecoraro et al., situation but takes the cumulative precipitation measured on the day of
2019; Stanley et al., 2020) and numerically estimate occurrences of the landslide occurrence plus the day before it (Peruccacci et al., 2017).
landslides by discriminating locations that previously experienced The performance is stored once again and the procedure continues
landsliding from those that did not fail (Bragagnolo et al., 2020; Guo backward, taking each time the cumulative precipitation value for a
et al., 2022). Overall, LEWSs issue landslide alerts solving a classifica­ fixed window and checking whether its scalar use would lead to a per­
tion task, and specifically by estimating the dependence between land­ formance increase (e.g., Afungang and Bateira, 2016; Wang et al., 2021).
slide presence/absence data and precipitation proxies. This is different The routine usually stops when the maximum accuracy is reached. This
from traditional susceptibility studies, which are solely based on static allows one to intuitively visualize the intensity-duration relationship
terrain and geological characteristics (Reichenbach et al., 2018; Titti (Mathew et al., 2014), as a means to understand how much precipitation
et al., 2022). The occurrence probability coming from LEWSs is there­ cumulated over time is most likely to trigger landslides for specific
fore mosty interpreted as the temporally-dynamic element of the hazard landscapes (Zhao et al., 2022).
definition (e.g., Ahmed et al., 2020; Steger et al., 2023). Landslide types and weather characteristics can largely influence the
At the core of any LEWSs is the notion of precipitation thresholds length of the cumulative time window. For instance, LEWSs for shallow
(Harilal et al., 2019; Rosi et al., 2021). These are essentially precipita­ landslides usually define the precipitation threshold over a shorter
tion values above which an alert is issued to local communities to amount of time (e.g., one week reported by Nikolopoulos et al., 2014) as
minimize the risk of being exposed to landsliding (Intrieri et al., 2012; compared to deep-seated failures (e.g., sixty days reported by Martelloni

* Corresponding author.
E-mail address: l.lombardo@utwente.nl (L. Lombardo).

https://doi.org/10.1016/j.envsoft.2023.105833
Received 11 August 2023; Received in revised form 18 September 2023; Accepted 20 September 2023
Available online 22 September 2023
1364-8152/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

et al., 2012). This is mainly because of the lag time between precipita­ resolution of the CHIRPS dataset constitutes the main criterion we use to
tion and failure. In other words, water infiltration could take longer for decide whether the spatial accuracy of a given reported landslide event
deep-seated landslides as compared to shallow ones and thus, requires a is suitable for our research. This means that if a reported landslide event
longer lag time to produce slope instabilities. However, this does not is identified with ~5 km accuracy, we include them in our database
mean that the influence of long-term precipitation could be ignored because it would still match the CHIRPS resolution. Any larger uncer­
while predicting shallow-seated landslides or vice versa for deep-seated tainty in the landslide geolocation leads to a rejection instead. The same
ones. Both short- and long-term precipitation records are useful to better spatial structure, therefore, dictated the choice of our mapping unit. In
understand conditions leading to the genesis of landslides. For instance, other words, we partitioned the Turkish landscape in a regular lattice of
prolonged dry periods followed by heavy rainfall could affect the approximately 5 km side.
shrinking-swelling response of clay-rich hillslopes and cause instabilities
(Schulz et al., 2018; Tichavský et al., 2019). This implies that aggre­ 2.2. Long Short-Term Memory networks
gating the precipitation discharged over time in LEWSs may not be the
ideal way to take into account this complex dynamic process. And yet, LSTM networks belong to a special class of recurrent neural networks
the geoscientific community has mostly worked with single scalar that can address vanishing gradient problems (Hochreiter, 1998) and
rainfall values so far. have been successfully employed in sequential data modeling (Lipton
Using a single aggregated measure of precipitation over time ne­ et al., 2016). Fig. 1 summarizes the structure of our LSTM model.
glects the potential information a data-driven model can gain from using LSTM uses three key gate functions to control the information flow
the entire precipitation time series. This is actually the main motivation process (Hochreiter and Schmidhuber, 1997), including the input gate
of this manuscript. We question whether we could pass the continuous ft , forget gate it , and output gate ot .
precipitation signal carrying rich spatiotemporal information to a binary Let x = {x1 , x2 , ..., xN } be a precipitation time series input. Wx , Wh ,
classifier to perform the very same task that traditional LEWSs solve. and b denote the weight of input, the weight of the hidden state, and the
This paper aims at proposing a novel approach for the prediction of bias, respectively. The forget gate ft determines whether the previous
precipitation-triggered landslides as a baseline that could lead to a new information is to be remembered or can be forgotten in the current time
LEWS protocol. Our framework explore a third dimension, time and to step t. It can be denoted as flows:
do so, we exploit the entire precipitation time series as an explanatory ( )
variable of our predictive model. ft = σ Wfx xt + Wfh ht− 1 + bf (1)
We use speech recognition architectures based on a deep learning
algorithm known as Long Short-Term Memory networks (LSTM, Zhou where σ is the sigmoid activation function, and ht− 1 is the previous
et al., 2016). These networks are capable of solving a classification task hidden memory state. The input gate it determines the input information
by using time series or functional data instead of a scalar information. updating and ̃ct memorizes the new information, which is denoted as
Their invention has been originally proposed in the context of text follows:
reading (Cheng et al., 2016) or speech recognition (Sak et al., 2014). In it = σ(Wix xt + Wih ht− 1 + bi ) (2)
the latter applications, they can discriminate two or multiple speakers as
a function of the time series of vocal cord vibrations, recognizing unique ct = tan h(Wcx xt + Wch ht− 1 + bc )
̃ (3)
voices, tones, and other characteristics. Similarly, we envision the use of
speech recognition to discriminate landslide presences/absences in where tan h is the activation function. The new memory cell ct the state
space and time, as a function of full precipitation time series. To test this is then updated as follows:
assumption, we go through a large number of sources reporting land­
slides including technical reports, newspapers, and scientific articles, ct = ft ⊙ ct− 1 + it ⊙ ̃ct (4)
and gather a precipitation-trigger landslide catalog of Turkey covering
where ct− 1 is the previous memory cell state and ⊙ is the Hadamard
the period between 2008 and 2020. We use this dataset to test our
product. Finally, the output gate ot controls the output activation, and
speech recognition idea in the context of landslide prediction.
the hidden memory state sent to the next time step is defined as follows:

2. Data and methods ot = σ (Wox xt + Woh ht− 1 + bo ) (5)

2.1. Landslide database ht = ot ⊙ tan h(ct ) (6)

We generate a landslide database covering the period between 2008 In the final time step, we apply a softmax activation function on the
and 2020. To do so, we scan through numerous sources including na­ hidden state hN , and regard the maximum score in the output as the final
tional, local printed and digital media reports (e.g., https://egazete. prediction.
cumhuriyet.com.tr/yayinlar) as well as research papers, incidents re­
ports from government and emergency agencies (e.g., https://www.afad 2.3. Model implementation
.gov.tr/afet-analiz and https://katalog.devletarsivleri.gov.tr). We
search for several keywords in Turkish (e.g., landslide, slide, mass We use the mapped landslide inventory to build the binary classifi­
movement, slump) to identify landslide occurrences and their occur­ cation model. This model boils down to a space-time occurrence prob­
rence dates. For each reported landslide event, we also check the ability estimator whose association between landslide presence/absence
available optical images through Google Earth and/or PlanetScope (3–5 data and precipitation makes use of the full precipitation time series. In
m), Rapid Eye (5 m) images acquired from Planet Labs (Planet Team, other words, if traditional landslide early warnings are based on a scalar
2017). This is done to confirm the spatial and temporal accuracy of the measure of precipitation (the sum over a fixed time window), our model
information for each event before adding it to the landslide database we does not summarize the time series in a single value but rather makes use
generate for this research. The final output of the procedure is a of the raw and full temporal signal instead. The functional representa­
point-based landslide database, whose geolocation is associated with the tion of the precipitation signal is extracted from CHIRPS data.
day of occurrence. To do so, we first extract daily precipitation values from the day of
For the specific task of this experiment, we couple landslide obser­ landslide occurrence to the 60th day before the event. We then obtain
vations with Climate Hazards Group InfraRed Precipitation with Station precipitation time series (thus made of 60 sequential time-points) for all
Data (CHIRPS, spatial resolution is 0.05◦ ; Funk et al., 2015). The spatial landslide locations in the inventory and replicate the same procedure for

2
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

Fig. 1. Schematic diagram summarizing the structure of our LSTM model.

landslide-absence locations. 2.4. Model benchmark

To create a reliable presence dataset, we remove any landslide point
for which the rainfall was zero. Taking away presence data with no We also compare our approach against an alternative structure
precipitation on the day the slope failure ensures removing any potential typical of traditional Early Warning Systems (Segoni et al., 2018). These
noise due to either inaccurate landslide reports or precipitation data. As are universally based on an intensity-duration relationship, reflected in
for the absences, the situation requires a number of additional steps. For the model by cumulating the rainfall amount over a given time window
instance, the number of potential absences to sample from the whole and thus obtaining a single scalar value. In doing so, we allow our
Turkish territory and repeated on a daily basis for 13 years, is extremely version of LEWS to be validated against a consolidated benchmark,
large. We thus start calculating the mean and standard deviation of the comparing the effect of using the whole time series and a single repre­
landslide-presence precipitation time series and set the mean plus 2 sentative value of the same. The performance metrics (AUC, F1-Score,
standard deviations as the minimum threshold to filter in landslide- and Cohen’s Kappa coefficient) mentioned above are also used in this
absence samples. We only retain landslide-absence samples as those case.
that have at least one extreme precipitation day (i.e., at least one day
above the μ+2σ obtained from the presences) across the 60-day time 3. Results
series. This operation is included to remove any trivial information. In
other words, if we would not include this criterion, the most common The landslide inventory we put together featured 2380 individual
situation in an arid territory such as the Turkish one, would potentially precipitation-triggered landslides that occurred between 2008 and 2020
extract time series where no rainfall occurred for large time windows. in Turkey (Fig. 2). Looking at their co-location within the 5 km lattice
Such information is not what we seek to model, as it should be obvious defined by the CHIRPS data, and after filtering out dry occurrence dates,
that no precipitation-induced landslide can take place without precipi­ the resulting presence database was reduced to 680 landslide-presence
tation in the first place. To avoid simply learning the difference between grids. The absences were then generated by selecting a balanced and
dry and wet conditions, extracting rainfall time series with at least one random sample from the remaining 5 km grids that fulfilled the criterion
extreme would challenge the model to discriminate the influence of of hosting at least an extremely rainy day. We would like to stress here
different precipitation regimes. that such data construction was also iteratively repeated to obtain a
The network structure and hyperparameters are important for deep representative sample of the suitable absence instances across the whole
learning-based modeling. In this study, the model architecture has a space-time domain under consideration.
LSTM hidden layer, further containing 16 hidden units. The model fol­ We then merged the presences with the random subsets of absence
lows an output layer with two neurons, each one equipped with a soft­ instances to build the LSTM networks. Fig. 3 shows a graphical example
max function. During the optimization, we use the categorical cross- of the precipitation time series associated with randomly selected
entropy as the loss function and ADAM is the optimizer of our choice landslide and no-landslide locations in 2020. This is something we show
(e.g., Kingma, 2014; Li et al., 2020). We set the batch size and epoch as to highlight how complex the differentiation between the two classes is
64 and 30, respectively (Brownlee, 2018). We implemented the LSTM if done through visual comparison. In fact, no evident difference
model using Python language under the Keras framework (Gulli and Pal, emerges between the two classes. It is for this reason that a neural
2017). network architecture is particularly appealing as it is capable of differ­
To assess the model prediction performance, we use three different entiating between the two sets of presence/absence time series (Karim
performance estimators: (1) the receiver operating characteristic curve et al., 2018).
(AUC, Fawcett, 2006), (2) F1-score (Singhal, 2001), and (3) Cohen’s We tested the sensitivity of our model using AUC, F1-score, and
Kappa coefficient (Kraemer, 2014). Ultimately, we apply a 10-fold Cohen’s Kappa coefficient, these being computed from two separate
cross-validation procedure to validate our model based on the experiments. The first one revolves around iteratively combining 10
balanced dataset. random batches of 680 absences from a large database, together with the
680 presences. Each one of these ten newly formed datasets will then
undergo the same LSTM procedure, whose performance is assessedvia a
random10-fold cross-validation. This operation ensures testing for the

3
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

Fig. 2. Spatial distribution of landslides occurred in Turkey between 2008 and 2020.

Fig. 3. Randomly selected precipitation time series for landslide-presence and landslide-absence conditions in 2020. The red and blue lines are just two examples we
picked to highlight the similarity between the two datasets. Even the absence time series host extreme rainfall discharges comparable to those shown in the presence
counterpart.

model robustness varying the representative precipitation regimes over series. The same panel shows all three performance metrics together,
the Turkish landscape and also over time. The second experiment re­ where an incremental performance is achieved essentially across all of
peats the same operation but shrinks the time window that defined the them at increasing lengths of the considered time windows.
rainfall time series, from an initial length of 60 days–10 days for a total We consider this a reasonable result because some additional infor­
of 6 nested tests (60, 50, 40, 30, 20, 10 days). mation may be contained over larger periods. But, it should also be
Results in Fig. 4 show that the AUC, F1-score, and Cohen’s s coeffi­ stressed that the power of using the whole time series is already visible
cient respectively range between 0.86 and 0.96, 0.83–0.92, and simply within 10 days, where the performances are particularly high
0.63–0.80, considering all experiments described above. The AUC is a nonetheless.
performance diagnostic commonly employed for classification purposes. For any new experimental design and associated numerical solution
For this reason, its values have been classified to summarize the results to be valid, one has to benchmark the results against a scientific stan­
into acceptable (0.7 < AUC <0.8), excellent (0.8 < AUC <0.9), and dard. In our case, this was done by keeping the LSTM architecture but
outstanding (0.9 < AUC <1), with 0.5 and 1.0 respectively defining the exploring the variation in performance obtained by using cumulated
purely random and perfect classification cases (see, Hosmer and Leme­ precipitation, as per Guzzetti et al. (2020), instead of the whole time
show, 2000). For this reason, an overall mean AUC of 0.920 across all series. The results are shown in Fig. 5b, c, and 5d where a drastic drop
tests indicates an outstanding performance of our proposed approach stands out across the whole spectrum of performance diagnostics.
that relied on the full precipitation time series. As for the best and worst Looking at the AUC values estimated for different time windows, the
models (see Fig. 5a), what stands out is that using the 60-day precipi­ results obtained with a scalar representation of the precipitation signal
tation time series produces an AUC of 0.934, whereas achieving the appear very far from those estimated with our full signal approach. In
lowest AUC of 0.906 is associated with the 20-day precipitation time fact, the maximum AUC (approx. 0.75) in the scalar case is reached at a

4
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

Fig. 4. Model prediction performance for different sampling schemes. Each boxplot shows the range of AUC based on 10-fold cross-validation in a given length of
precipitation time series and a given random seed of absence samples selection. The red points in the boxplot refer to the mean value, whereas the red box indicates
the limits of lower and upper quartiles calculated using 60-day precipitation.

cumulated rainfall of 50 days. This almost corresponds to a 20% drop in probability is reached for the single 5 km × 5 km grid where the failure
performance with respect to the maximum AUC reached for a 60-day took place. Another interesting element to be examined in Fig. 6 cor­
time window, when using the entire precipitation series. The very responds to the right panel at the bottom. There, for the unstable grid
same pattern is retrieved for the remaining two performance indicators. mentioned above, we plot the sequence of daily susceptibility values,
Having demonstrated the gain in predictive power, we opted to together with the rainfall time series. Despite the model takes into ac­
include a hindcast example, converting the results to map form. To do count 60 days of precipitation, it is still sensitive to impulsive rainfall
so, we focused on the North-Eastern sector of Turkey, where rainfall is discharges, with trends in the probability that follow potential spikes in
frequently responsible for landslide occurrences, even fatal ones (Görüm precipitation discharges. This is quite typical of the LSTM architecture as
and Fidan, 2021). There, we simulated the daily dynamism in the sus­ it progressively “forgets” signals from the past and assigns larger weights
ceptibility patterns for June 2020, following our approach and specif­ as close as possible to the date of interest. This characteristic is quite
ically for the model defined over the 60-day time series (because it suitable in the landslide context because as a community, we usually
proved to be numerically the best). The occurrence of a reference discriminate between preparatory and triggering rainfall (Mondini
landslide in the area is marked on the 13th of June in our inventory. This et al., 2023; Steger et al., 2023).
is also the day in our simulation when the maximum occurrence

5
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

Fig. 5. Panel (a) reports the three performance metrics computed for the LSTM architecture fed with the continuous precipitation signal. Panels (b), (c), and (d) show
the same information (black lines) but overlaid to a 2-D space where the same metrics are obtained for an LSTM fed with a scalar representation of the precipitation
(orange lines).

Fig. 6. A hindcast example demonstrating predicted landslide probably varying over time with respect to precipitation time series.

4. Discussion that the idea of using speech recognition in landslide predictive studies
may constitute the foundation for next-generation early warning sys­
The proposed approach returned outstanding performances across tems in the future. However, this method still requires additional tests
the whole spectrum of tests and model diagnostics, whether we included and below we will share our vision of what still needs to be done before
random tests performed by changing the absence instances, or for all it may become an operational tool.
tests run at different time windows. Overall, this is a positive indication The most intuitive element to be added corresponds to the use of

6
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

terrain characteristics in the model. The current model we propose stems for predicting precipitation-induced landslides. Specifically, as an
from the traditional setting where rainfall intensity and duration are alternative method to the traditional accumulated precipitation proxies,
responsible for the landslide occurrence probability, without featuring here we explore the use of a Neural Network architecture typical of
landscape characteristics typical of landslide susceptibility studies. speech recognition. This offers the ability to pass to the model a func­
However, time, technological advances and data availability have tional representation of the precipitation signal rather than a scalar
reached a level of maturity where space-time models could finally approximation of it. To support this complex experimental design, we
feature all this information at once. An example already exists where developed a completely new landslide inventory exclusively reporting
terrain and rainfall characteristics (albeit in a scalar form) are simulta­ precipitation-induced landslides, that occurred between 2008 and 2022
neously regressed against dynamic landslide occurrence data (Stanley across Turkey.
et al., 2021; Steger et al., 2023). The performance we retrieve appears to be significantly better than
Another important aspect to be improved could involve moving the traditional counterpart (i.e., 20% higher prediction power). There­
away from a coarse mapping unit. The 5 km mapping unit we chose fore, our results indicate that the full-time series may carry more useful
could be further downscaled to provide valuable information for information when classifying locations likely to undergo landsliding or
regional or even catchment-scale applications. However, to do so, not. Despite the achieved results, many other tests are required before a
satellite-based precipitation may not be sufficient as they are mostly speech-recognition could actually make its way into landslide early
coarse and affected by a number of potential biases and uncertainties. warning systems. Among these tests, other study areas are certainly
One possible solution may involve the use of weather stations equipped required, but the most important element to be addressed lies in the use
with radar sensors. This would certainly ensure much finer spatial and of forecasted rather than hindcasted precipitation. If such data would
temporal resolutions. Notably, radar data constitutes the backbone of also confirm its success over the traditional scalar version of the same, a
many state-of-the-art weather forecast systems. In its current state, our new generation of alert system may be developed, based on this inno­
speech-recognition approach is exclusively relying on CHIRPS and spe­ vative speech-recognition idea.
cifically on past precipitation estimates. Therefore, it can only be used
for hindcasting purposes. This is something that still needs to be vali­ Declaration of competing interest
dated and compared against other analogous products. A suitable
candidate in this sense may be the MSWEP V2 (Beck et al., 2019) as it The authors declare the following financial interests/personal re­
both brings high spatial and temporal details, with a level of accuracy lationships which may be considered as potential competing interests:
proven to be better than most satellite-based alternative products. Luigi Lombardo reports financial support was provided by National
However, aside from the hindcasting aspects, we envision future efforts Natural Science Foundation of China.
to expand beyond the limitations of looking backward in time and rather
test it for nowcast and forecast applications. This would require Data availability
re-training the LSTM architecture on equally nowcasted and forecasted
precipitation measures. Ground-based radar could certainly offer a po­ https://doi.org/10.5281/zenodo.8103276
tential solution, but this is also valid for other space-born satellite mis­
sions. For instance, the model-based Goddard Earth Observing Acknowledgement
System-Forward Processing (GEOS-FP) precipitation (Molod et al.,
2012; Suarez et al., 2008) offers a forecast which has already been This work was supported by the Joint Funds of the National Natural
exploited as a tool to develop an early warning system for Science Foundation of China (U21A2013) and the Fundamental
precipitation-induced landslides (Khan et al., 2022). The use of such Research Funds for National Universities, China University of Geo­
precipitation forecast instead of CHIRPS may open up a LEWS tool. This sciences (Wuhan). Initial idea proposed by Ashok Dahal. Experimental
of course brings us back to the spatial resolution issue, albeit this would design developed by Zhice Fang, Luigi Lombardo and Hakan Tanyas.
come with an increase in the temporal details. The topic of temporal The same authors have prepared scientific illustrations and written the
resolution is something particularly interesting, especially because of manuscript. Zhice Fang has also worked on the implementation. Tolga
the speech-recognition architecture we propose. In fact, the current Gorum provided the data. Yi Wang secured the above mentioned funds.
model scans through the precipitation signal expressed daily. Therefore,
it could be scientifically interesting to see whether adding sub-daily References
information (3-hourly, hourly, or even sub-hourly) would lead to
different results and explore why. Going back to the topic related to the Afungang, R.N., Bateira, C.V., 2016. Temporal probability analysis of landslides
uncertainty in the precipitation products, better accuracy could also lead triggered by intense rainfall in the Bamenda Mountain Region, Cameroon. Environ.
Earth Sci. 75, 1032. https://doi.org/10.1007/s12665-016-5835-7.
to variation in the model output. For instance, in the current settings, we Ahmed, B., Rahman, M.S., Sammonds, P., Islam, R., Uddin, K., 2020. Application of
had to remove part of the landslide presence information because some geospatial technologies in developing a dynamic landslide early warning system in a
instances reported zero or negligible precipitation on the day of the humanitarian context: the Rohingya refugee crisis in Cox’s Bazar, Bangladesh.
Geomatics, Nat. Hazards Risk 11, 446–468. https://doi.org/10.1080/
landslide occurrence. Relying on better precipitation products, even 19475705.2020.1730988.
from interpolated discharges measured over a rain-gauge network could Beck, H.E., Wood, E.F., Pan, M., Fisher, C.K., Miralles, D.G., Van Dijk, A.I., McVicar, T.R.,
potentially retain useful slope instability information. Ultimately, Adler, R.F., 2019. MSWEP V2 global 3-hourly 0.1 precipitation: methodology and
quantitative assessment. Bull. Am. Meteorol. Soc. 100 (3), 473–500.
another potential improvement we envision relies on the use of a clas­
Bragagnolo, L., da Silva, R.V., Grzybowski, J.M.V., 2020. Landslide Susceptibility
sified landslide inventory. Currently, we used a national-level landslide Mapping with R. Landslide: A Free Open-Source GIS-Integrated Tool Based on
information and as such, it did not report which landslide type each data Artificial Neural Networks, vol. 123. Environmental Modelling & Software, 104565.
Brownlee, J., 2018. What is the difference between a batch and an epoch in a neural
point belonged to. However, at least theoretically, one should expect
network. Mach. Learn. Mastery 20.
that different landslide types should behave differently form a pure Cheng, J., Dong, L., Lapata, M., 2016. Long Short-Term Memory-Networks for Machine
hydrological perspective. Therefore, by testing the current modeling Reading, 06733 arXiv Prepr. arXiv1601.
architecture on specific landslide types should lead to even better results Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recogn. Lett. 27, 861–874.
https://doi.org/10.1016/j.patrec.2005.10.010.
as compared to those obtained in this work. Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G.,
Rowland, J., Harrison, L., Hoell, A., Michaelsen, J., 2015. The climate hazards
5. Conclusions infrared precipitation with stations—a new environmental record for monitoring
extremes. Sci. Data 2, 150066. https://doi.org/10.1038/sdata.2015.66.
Gariano, S.L., Guzzetti, F., 2016. Landslides in a changing climate. Earth Sci. Rev. 162,
This research offers a different take on the typical modeling structure 227–252. https://doi.org/10.1016/j.earscirev.2016.08.011.

7
Z. Fang et al. Environmental Modelling and Software 170 (2023) 105833

Görüm, T., Fidan, S., 2021. Spatiotemporal variations of fatal landslides in Turkey. Function Network. Landslides 7, 325–338. https://doi.org/10.1007/s10346-010-
Landslides. https://doi.org/10.1007/s10346-020-01580-7. 0229-5.
Gulli, A., Pal, S., 2017. Deep Learning with Keras. Packt Publishing Ltd. Pecoraro, G., Calvello, M., Piciullo, L., 2019. Monitoring strategies for local landslide
Guo, Z., Torra, O., Hürlimann, M., Abancó, C., Medina, V., 2022. FSLAM: a QGIS plugin early warning systems. Landslides 16, 213–231. https://doi.org/10.1007/s10346-
for fast regional susceptibility assessment of rainfall-induced landslides. Environ. 018-1068-z.
Model. Software 150, 105354. Peruccacci, S., Brunetti, M.T., Gariano, S.L., Melillo, M., Rossi, M., Guzzetti, F., 2017.
Guzzetti, F., Gariano, S.L., Peruccacci, S., Brunetti, M.T., Marchesini, I., Rossi, M., Rainfall thresholds for possible landslide occurrence in Italy. Geomorphology 290,
Melillo, M., 2020. Geographical landslide early warning systems. Earth Sci. Rev. 200 39–57. https://doi.org/10.1016/j.geomorph.2017.03.031.
https://doi.org/10.1016/j.earscirev.2019.102973. Planet Team, 2017. Planet application program interface. San Francisco, CA. In: Space
Harilal, G.T., Madhu, D., Ramesh, M.V., Pullarkatt, D., 2019. Towards establishing for Life on Earth. https://api.planet.com [WWW Document].
rainfall thresholds for a real-time landslide early warning system in Sikkim, India. Reichenbach, P., Rossi, M., Malamud, B.D., Mihir, M., Guzzetti, F., 2018. A review of
Landslides 16, 2395–2408. https://doi.org/10.1007/s10346-019-01244-1. statistically-based landslide susceptibility models. Earth Sci. Rev. 180, 60–91.
Hochreiter, S., 1998. The vanishing gradient problem during learning recurrent neural https://doi.org/10.1016/j.earscirev.2018.03.001.
nets and problem solutions. Int. J. Uncertain. Fuzziness Knowledge-Based Syst. Rosi, A., Segoni, S., Canavesi, V., Monni, A., Gallucci, A., Casagli, N., 2021. Definition of
107–116. https://doi.org/10.1142/S0218488598000094, 06. 3D rainfall thresholds to increase operative landslide early warning system
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9, performances. Landslides 18, 1045–1057. https://doi.org/10.1007/s10346-020-
1735–1780. 01523-2.
Hosmer, D.W., Lemeshow, S., 2000. Applied Logistic Regression. John Wiley & Sons, Sak, H., Senior, A., Beaufays, F., 2014. Long Short-Term Memory Based Recurrent Neural
New York. Network Architectures for Large Vocabulary Speech Recognition arXiv Prepr.
Intrieri, E., Gigli, G., Mugnai, F., Fanti, R., Casagli, N., 2012. Design and implementation arXiv1402.1128.
of a landslide early warning system. Eng. Geol. 147–148, 124–136. https://doi.org/ Schulz, W.H., Smith, J.B., Wang, G., Jiang, Y., Roering, J.J., 2018. Clayey landslide
10.1016/j.enggeo.2012.07.017. initiation and acceleration strongly modulated by soil swelling. Geophys. Res. Lett.
Karim, F., Majumdar, S., Darabi, H., Chen, S., 2018. LSTM fully convolutional networks 45, 1888–1896. https://doi.org/10.1002/2017GL076807.
for time series classification. IEEE Access 6, 1662–1669. https://doi.org/10.1109/ Segoni, S., Piciullo, L., Gariano, S.L., 2018. A review of the recent literature on rainfall
ACCESS.2017.2779939. thresholds for landslide occurrence. Landslides 15, 1483–1501. https://doi.org/
Khan, S., Kirschbaum, D.B., Stanley, T.A., Amatya, P.M., Emberson, R.A., 2022. Global 10.1007/s10346-018-0966-4.
landslide forecasting system for hazard assessment and situational awareness. Front. Singhal, A., 2001. Modern information retrieval: a brief overview. IEEE Data Eng. Bull.
Earth Sci. 24, 35–43.
Kingma, D.P., 2014. A Method for Stochastic Optimization. ArXiv Prepr. Stanley, T.A., Kirschbaum, D.B., Sobieszczyk, S., Jasinski, M.F., Borak, J.S., Slaughter, S.
Kraemer, H.C., 2014. Kappa Coefficient. Wiley StatsRef Stat. Ref online 1–4. L., 2020. Building a Landslide Hazard Indicator with Machine Learning and Land
Li, S., Chen, H., Wang, M., Heidari, A.A., Mirjalili, S., 2020. Slime mould algorithm: a Surface Models, vol. 129. Environmental Modelling & Software, 104692.
new method for stochastic optimization. Future Generat. Comput. Syst. 111, Stanley, T.A., Kirschbaum, D.B., Benz, G., Emberson, R.A., Amatya, P.M., Medwedeff, W.,
300–323. https://doi.org/10.1016/j.future.2020.03.055. Clark, M.K., 2021. Data-driven landslide nowcasting at the global scale. Front. Earth
Lipton, Z.C., Kale, D.C., Elkan, C., Wetzel, R.C., 2016. Learning to Diagnose with LSTM Sci.
Recurrent Neural Networks. BT - 4th International Conference on Learning Steger, S., Moreno, M., Crespi, A., Zellner, P.J., Gariano, S.L., Brunetti, M.T., Melillo, M.,
Representations. Conference Track Proceedings. ICLR 2016, San Juan, Puerto Rico, Peruccacci, S., Marra, F., Kohrs, R., Goetz, J., Mair, V., Pittore, M., 2023.
May 2-4, 2016. Deciphering seasonal effects of triggering and preparatory precipitation for
Martelloni, G., Segoni, S., Fanti, R., Catani, F., 2012. Rainfall thresholds for the improved shallow landslide prediction using generalized additive mixed models.
forecasting of landslide occurrence at regional scale. Landslides 9, 485–495. https:// Nat. Hazards Earth Syst. Sci. 23, 1483–1506. https://doi.org/10.5194/nhess-23-
doi.org/10.1007/s10346-011-0308-2. 1483-2023.
Mathew, J., Babu, D.G., Kundu, S., Kumar, K.V., Pant, C.C., 2014. Integrating Suarez, M.J., Rienecker, M.M., Todling, R., Bacmeister, J., Takacs, L., Liu, H.C., Gu, W.,
intensity–duration-based rainfall threshold and antecedent rainfall-based probability Sienkiewicz, M., Koster, R.D., Gelaro, R., 2008. The GEOS-5 Data Assimilation
estimate towards generating early warning for rainfall-induced landslides in parts of System-Documentation of Versions 5, 0. 1, 5.1. 0, and 5.2. 0.
the Garhwal Himalaya, India. Landslides 11, 575–588. https://doi.org/10.1007/ Tichavský, R., Ballesteros-Cánovas, J.A., Šilhán, K., Tolasz, R., Stoffel, M., 2019. Dry
s10346-013-0408-2. spells and extreme precipitation are the main trigger of landslides in central Europe.
Melillo, M., Brunetti, M.T., Peruccacci, S., Gariano, S.L., Roccati, A., Guzzetti, F., 2018. Sci. Rep. 9, 14560 https://doi.org/10.1038/s41598-019-51148-2.
A Tool for the Automatic Calculation of Rainfall Thresholds for Landslide Titti, G., Napoli, G.N., Conoscenti, C., Lombardo, L., 2022. Cloud-based interactive
Occurrence, vol. 105. Environmental Modelling & Software, pp. 230–243. susceptibility modeling of gully erosion in Google Earth Engine. Int. J. Appl. Earth
Molod, A., Takacs, L., Suarez, M., Bacmeister, J., Song, I.-S., Eichmann, A., 2012. The Obs. Geoinf. 115, 103089.
GEOS-5 Atmospheric General Circulation Model: Mean Climate and Development Wang, N., Lombardo, L., Gariano, S.L., Cheng, W., Liu, C., Xiong, J., Wang, R., 2021.
from MERRA to Fortuna. Using satellite rainfall products to assess the triggering conditions for hydro-
Mondini, A.C., Guzzetti, F., Melillo, M., 2023. Deep learning forecast of rainfall-induced morphological processes in different geomorphological settings in China. Int. J.
shallow landslides. Nat. Commun. 14, 2466. https://doi.org/10.1038/s41467-023- Appl. Earth Obs. Geoinf. 102, 102350 https://doi.org/10.1016/j.jag.2021.102350.
38135-y. Zhao, L., Liu, M., Song, Z., Wang, S., Zhao, Z., Zuo, S., 2022. Regional-scale modeling of
Nikolopoulos, E.I., Crema, S., Marchi, L., Marra, F., Guzzetti, F., Borga, M., 2014. Impact rainfall-induced landslides under random rainfall patterns. Environ. Model. Software
of uncertainty in rainfall estimation on the identification of rainfall thresholds for 155, 105454.
debris flow occurrence. Geomorphology 221, 286–297. https://doi.org/10.1016/j. Zhou, P., Shi, W., Tian, J., Qi, Z., Li, B., Hao, H., Xu, B., 2016. Attention-based
geomorph.2014.06.015. bidirectional long short-term memory networks for relation classification. In:
Osanai, N., Shimizu, T., Kuramoto, K., Kojima, S., Noro, T., 2010. Japanese early- Proceedings of the 54th Annual Meeting of the Association for Computational
warning for debris flows and slope failures using rainfall indices with Radial Basis Linguistics, 2, pp. 207–212. Short Papers).