Journal of Cleaner Production 314 (2021) 128073

Contents lists available at ScienceDirect

Journal of Cleaner Production

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

Constructing the machine learning techniques based spatial drought

vulnerability index in Karnataka state of India
Sunil Saha a, *, Priyanka Gogoi a, Amiya Gayen b, Gopal Chandra Paul a
a
Department of Geography, University of Gour Banga, Malda, West Bengal, India
b
Department of Geography, University of Calcutta, Kolkata, West Bengal, India

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

Handling Editor: Bin Chen The drought induced vulnerability is owed much to rapid modernization, climate extremes, and over exploitation
of natural resources. However, a natural phenomenon, the drought is amplified by anthropogenic activities that
Keywords: in its enormity influences water availability, agricultural productivity, ecosystem, groundwater storage. The
Resilience Karnataka state of India is frequently affected by the drought that causes huge loss in agricultural sector and
Sensitivity
other allied sectors. Therefore, it is essential to measure the vulnerability status for the better management of
Exposure
natural resources in the state of Karnataka. No advanced models are being used yet to portray the drought
Relative drought vulnerability
Machine learning algorithms vulnerability status. Different advanced machine learning models are effective in predicting various physical
vulnerabilities. The aim of this study was to use sophisticated machine learning models to precisely define
relative drought vulnerability. In that endeavour, it used two advanced machine-learning algorithms (MLAs),
namely, Bagging and Artificial Neural Network (ANN) which are still not used in this field. Twenty-six meteo­
rological and socio-economical parameters were considered to find the most drought vulnerable areas. The
predisposing parameters were classified as resilience (7 parameters), sensitivity (9 parameters), and exposure (10
parameters). The researchers have produced drought vulnerability maps for overall condition, resilience,
sensitivity, and exposure. The relative drought vulnerability maps (RDVMs) clearly show that 40.87%–52.03% of
areas fall under very high vulnerability, which is situated in the central and eastern parts of the state. The
prediction capacity of newly built models was judged with efficiency, root mean square error (RMSE), true skill
statistics (TSS), Friedman and Wilcoxon rank test, and area under the curve (AUC) of receiver operating char­
acteristic (ROC). All of them showed satisfactory results - the RMSE value of 0.32 and 0.33, TSS values of 0.82
and 0.81, and AUC values of 86.50% and 84.20% as obtained by ANN and bagging models, respectively. The
produced RDVMs demonstrate the urgency of policy interventions to minimize vulnerability in prioritized areas.

1. Introduction categorize it as meteorological, socioeconomic, agricultural, and hy­

drological drought. Meteorological drought is triggered by the pattern of
The growing interest in developing methodologies for assessment of dry weather in an area; hydrological drought arises when the avail­
the vulnerability of drought has become an important research area ability of water decreases; agricultural drought is a situation in which
(Turner et al., 2003; Fraser, 2007) pertinently for policy prescription crops are impaired and socioeconomic drought arises when drought hits
and adjustment with climate change. Drought is different from aridity as people and their economies (Pandey et al., 2010; Thomas et al., 2016).
the former is temporary and the latter is permanent. It can be recurrent In the cases of Jharkhand, Chhattisgarh, and Orissa, nearly 13 million
and is variable in nature with regard to its spatial span, duration, in­ population fall below the poverty line only because of drought-induced
tensity, magnitude, and dryness quotient. Defining drought is difficult as agricultural income loss (Pandey and Bhandari, 2006).
the event is a set of complicated phenomena. It is a condition of insuf­ Droughts are inevitable in the contemporary episodes of climate
ficiency of water or in other words a condition of water shortage (Wil­ change. However, it may be monitored and predicted to reduce its
hite and Glantz, 1985; Dracup et al., 1980). The meteorologists adverse effect (Agnew, 1990; Smakhtin and Hughes, 2007). So, drought

* Corresponding author.
E-mail addresses: [email protected] (S. Saha), [email protected] (P. Gogoi), [email protected] (A. Gayen), gopalpaul.0321@gmail.
com (G.C. Paul).

https://doi.org/10.1016/j.jclepro.2021.128073
Received 11 January 2021; Received in revised form 18 June 2021; Accepted 21 June 2021
Available online 23 June 2021
0959-6526/© 2021 Elsevier Ltd. All rights reserved.
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

vulnerability studies are very useful as it helps in the impact assessment model is taken to build relative drought vulnerability map (RDVM), as it
of drought and the related phenomena. Though the definition of fabricates the decision tree with the help of each produced subsets, and
vulnerability is complex and varies with the purpose and field of ultimately assembles them within the final model (Hong et al., 2020).
research (Engström et al., 2020) the frangibility of a system to various Also, the ANN model can predict accurately through its innovative
disasters can be termed as its vulnerability (Kelly and Adger, 2000; IPCC, classification scheme with high generalization capability (Singha et al.,
2001). Vulnerability to drought can be measured using an exposure of 2020).
the people, sensitivity to outside threats, and resilience of that system The primary objective of this work is to prepare a relative drought
(Engström et al., 2020; Xiaoqian et al., 2013). The drought, which re­ vulnerability map in Karnataka using various meteorological and so­
sults from precipitation variability, climate extremes, global warming, cioeconomic indicators. In the field of drought vulnerability mapping,
etc., poses a great threat to the global environment and society (IPCC this is a novel attempt wherein the MLAs are vigorously used in a robust
et al., 2007). It has been estimated that the global economic losses are manner.
higher for drought in comparison to other meteorological disasters
(Tánago et al., 2016). So, drought risk reduction is also associated with 2. Materials and methods
indirect and direct impacts by targeted risk reduction and adjustment
has become global importance, as manifested through the present global 2.1. Study area
framework and initiative, like Sendai framework for the Disaster Risk
Reduction 2015–2030, 2018/2019 UNCCD Drought Initiative, and the The state of Karnataka is the 7th largest and an important state for
GAR Special Report on Drought in 2020, etc. (Hagenlocher et al., 2019). Information Technology and Information Technology Enabled Services
Tropical countries like India mainly rely on monsoon showers to fulfil (IT/ITES) industry. It covers an area of 191,791 sq. km or nearly 5.83%
their water demand for agriculture and rest activities (Zhang et al., within the entire geographical area of India (Fig. 1). It is ranked second
2015; Pei et al., 2018). With the growing industrialization, moderniza­ in drought-prone area ranking, following Rajasthan. Within 31 districts,
tion, and population pressure, water demand is increasing day by day. 18 districts are highly drought prone. In this state consecutive drought
Despite receiving heavy monsoon showers, a considerable part of India occurred in 2001–02, 2002–03, and 2003–04 (Biradar and Sridhar,
is drought-affected, as the rainfall is not uniformly spread. Besides, the 2009). About 80% of the geographical area in Karnataka is drought
variability of the monsoon shower and the delay in its onset are some of prone (Reddy and Prabhu, 2016). Generally, the rainfall in the primary
the other factors. According to DEWS (Drought Early Warning System), stage of the monsoon season (June and July) shows an uneven variation
42% of India’s land is facing drought. Drought studies especially in India in the Karnataka state, specifically in the current drought events of 2009,
are of great concern, as 70% of its rural population is dependent on 2012, 2014, and 2015 (Vyas and Bhattacharya, 2020). The rainfall
agriculture, which again is heavily dependent on monsoon rainfall. In­ deficit was observed as 27%, 49%, and 20–50% in June 2009, 2012, and
dian agriculture is the backbone of the country’s economy. 2014, respectively (Pai and Bhan, 2015). About 80% of the total popu­
To its help, many researchers have tried to assess the drought lation is forced to leave agricultural activities and join the labour force,
vulnerability mapping by applying different approaches (Simelton et al., both outside villages and at relief sites for drought impact during
2009). For example, the analytical hierarchy process (Yuan et al., 2015; 2002–2003. The capital city of Karnataka is Bangalore, referred to as
Shabbir and Ahmad, 2016), fuzzy logic (Saha et al., 2021; Cheng and ‘India’s Silicon Valley’. This state extends from 11◦ 37′ to 18◦ 28′ latitude
Tao, 2010), Multi-Criteria decision making (MCDM) model. Several and 74◦ 6′ to 78◦ 35′ E longitude. It comprises 31 districts. The Western
experiments were done to determine the susceptibility of drought and its Ghat, coastal plains, and Deccan Plateau are the three geographical re­
effects by applying the Standard Perception Index (SPI) (Sönmez et al., gions of Karnataka and the physiographic divisions include Southern
2005; Lakshmi et al., 2020). One such research is conducted in the Karnataka Plateau, Northern Karnataka Plateau, Coastal Karnataka re­
United States, where demographic, climate and environmental variables gion, and Central Karnataka Plateau. The drainage networks of Karna­
are used to determine the vulnerability to drought in each state taka comprise Krishna, Cauvery, Godavari, South Penner, Palar, North
(Engström et al., 2020). Another such study is conducted over middle Pennar, and numerous West Flowing Rivers. Karnataka experiences 3
Inner Mongolia of China in which vulnerability of drought was assessed types of climate regions. It is also classified into three meteorological
based on Exposure, resilience, and sensitivity (Liu et al., 2013). Jafri and regions of south and north interior Karnataka and coastal Karnataka.
others in their site suitability analysis of runoff harvesting sites in The highest temperature in summer is 34 ◦ C, while the lowest temper­
Sub-Desert regions have used the Fuzzy membership method (Jafari ature in the winter session is nearly 16 ◦ C. Most of the areas in this state
Shalamzari et al., 2019). Nowadays these models are not widely applied have a low intensity of rainfall of 1028.5 mm with failure of the
by researchers, as they cannot handle multi-criteria decision problems. south-west monsoon giving birth to occurrences of drought (Rajendran,
Presently, machine learning algorithms (MLAs) have been widely 2001). In terms of population, Karnataka’s rank is 9th in India. The
used in various research fields like gully erosion, wetland habitat, flood, population density of Karnataka is 319 persons/sq. km and the total
human health, and landslide vulnerability assessment (Gayen et al., population has reached 61.13 million in the last census in 2011. The sex
2020; Hong et al., 2018a,b; Mandal and Pal, 2020; Pal and Paul, 2020; ratio of Karnataka is 973 which exceeds the average sex ratio (940) in
Saha and Saha, 2020). Several researchers have opinionated that the India. Karnataka has 3 software technology parks and 47 IT/ITES SEZs.
machine learning model has provided better prediction results than the Karnataka houses the 4th biggest technology cluster within the globe.
conventional statistical models. In addition, machine learning algo­
rithms can handle data with high dimensionality and provide more ac­ 2.2. Machine learning models for RDV map
curate results (Gayen et al., 2019, 2020, 2020; Saha et al., 2021). Based
on the accuracy of the MLA in the present study Accuracy of the Bagging The relative drought vulnerability (RDV) maps were developed by
and artificial neural network (ANN) models were used in the present applying two well-accepted machine learning algorithms following five
study. The MLAs are still not widely used in the field of drought key steps (Fig. 2).
vulnerability assessment. The main research questions are that how far
1st step. : predisposing factors selection: The RDV factors were taken
the present study area is vulnerable to drought condition and secondly
based on field experiences, literature survey, and data availability. The
are MLAs more capable than the conventional methods in producing
selected predisposing factors were classified into three sub-categories (i.
precise spatial drought vulnerability maps? The present study is focused
e., exposure, sensitivity, and resilience).
on the MLAs of Bagging and ANN models for addressing the aforemen­
tioned research questions. Machine learning models can handle the 2nd step. : Inventory map and thematic data layer preparation: Data
complex behaviour of parameters (Selbesoglu, 2020). The Bagging related to drought-affected villages and vulnerability related factors

2
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 1. Location map of the study area; (a) India (b) State of Karnataka.

Fig. 2. Flowchart illustrated the applied methodology of Relative Drought Vulnerability mapping of Karnataka.

were collected to predict spatial drought vulnerability. Identified the (Nageswara Rao et al., 2005; Biradar and Sridhar, 2009).
drought-affected areas through field surveys, Karnataka state reports,
3rd step. : Spatial data layers standardization: The applied spatial data
and existing literature about the socio-economic impact on drought

3
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Table 1
Showing the directionality of influence of indicators of exposure to drought.
Sl. Variables Logical consideration Directionality of influence to Source
no. drought

1 Annual rainfall More the rainfall less the vulnerability Indirect IMD
2 Rainfall trend The decreasing trend will indicate an intensification of the dry condition Indirect Calculated
and vice versa
3 Mean drought intensity of 12 months More the intensity (SPI ≤1.0) more the risk Indirect Calculated
4 Drought magnitude of 12 months More magnitude means more risk Indirect Calculated
5 Severe drought frequency of 12 months More frequency means more risk Direct Calculated
(%)
6 Extreme drought frequency of 12 months More frequency means more risk Direct Calculated
(%)
7 Recurrence interval of severe drought 12 More the recurrence interval less the risk Indirect Calculated
month
8 Recurrence interval of extreme drought 12 More the recurrence interval less the risk Indirect Calculated
month
9 Critical Rainfall More the critical rainfall more will be the risk Direct IMD
10 Wet-day frequency More the frequency more will be the risk Indirect IMD

layers were standardized using the Fuzzy Membership function. The 2.3.1.1. Standard Precipitation Index (SPI). The SPI was prepared by
different sets of spatial data layers were standardized according to their McKee et al., in 1993 for controlling the effects of drought events. This
motive of sub-indices. index was approved by the WMO (World Meteorological Organization).
The calculation of SPI needs only precipitation data (Edwards, 1997; Lin
4th step. : Preparation of RDV maps: To build the RDV maps both ANN
et al., 2020; Liu et al., 2021). It is used for quantifying the precipitation
and Bagging models were applied with the help of training data sets.
deficit for multiple time intervals i.e., 48, 24, 12, 6, 3, and SPI-1 months
After that, selected parameters were used to classify the groups for the
(McKee et al., 1993). Here, individual SPI was calculated in the R
assessment of individual (i.e., resilience, sensitivity, and exposure)
environment using the rainfall data and Eq. (1):
indexing. Finally, the relative drought vulnerability map was developed
using three indices. Xi − X
SPI = (Eq. 1)
5th step. : Model’s comparison and validation: The AUC-ROC, accu­ σ
racy, efficiency, RMSE, and MAE were used to validate the implemented where, Xi = Precipitation of the station, X ‾ = Mean precipitation. σ =
models. Also, Wilcoxon and Friedman’s rank tests were applied to assess Standard Deviation. For the present study, only the extreme and severe
the differences in the predictive results of the models. drought classes were used to define the severity of the occurrence of
drought in Karnataka. The classification of drought used (Table 2) is
based on SPI values of drought classification. In the study, the duration
2.3. Preparation of data layers
of SPI used is 12 months as it better reflects the long-term precipitation
condition and can be applied to determine the stream flows, reservoir
In the present study, a multi criterion-based RDV mapping approach
levels, and also groundwater levels (Table 3).
through machine learning models is used. Consequently, spatial map­
The extreme and severe drought frequencies were calculated in
ping of existing drought-affected areas is important to predict the
percentage using formulae. 2:
interrelationship between the distribution of drought-affected areas and
the influencing factors of the drought vulnerability. A total of 140 vil­ Ni
Ni,100 = × 100 (Eq. 2)
lages are identified. Out of which, 70% of villages are taken to develop i.n
RDVM, and the other 30% of village’s locations were employed to assess
the prediction capacity of the implemented models (Zabihi et al., 2019; where, Ni, 100 denotes the numbers of drought for a particular timeframe
Gayen et al., 2019; Saha et al., 2020). In general, the foremost challenge i in 100 years, Ni represents the number of months with droughts for a
was to define an effective set of indicators that are capable of fulfilling time scale i in the n year set, i = time scale (i.e., 6, 12 months). Drought
the objective. In that endeavour, various meteorological and socioeco­ magnitude is the cumulative water scarcity (CWS) during the drought
nomic parameters are selected based on previous literature, field expe­ period and means drought intensity (MID) is the average of this CWS
riences, and data availability, which are then classified into three during the drought period thus MD and MID can be calculated as (Saha
categories - resilience, sensitivity, and exposure (Saha et al., 2021). et al., 2021) using Eq. (3) and Eq. (4):
These data were collected from various sources mentioned in Tables 2, 4 ∑
MD = SPI ij (Eq. 3)
and 5. For mapping the RDV rainfall data were collected for the period of
1901–2019. Thereafter, different drought related parameters were MD
calculated from the collected rainfall data. MID = (Eq. 4)
m

2.3.1. Exposure indicators where, SPIij indicates the SPI values of a particular drought period, at j
These parameters determine how much of the area or its population timeframe, m is the number of months. For the calculation of recurrence
is exposed to drought. The indicators used for representing exposure interval (RI) at first, all the SPI values were arranged in ascending order
includes extreme drought frequency (%), severe drought frequency (%), and after that, all positive SPI values are assigned 0. Finally, the Cali­
mean drought intensity, drought magnitude, recurrence interval of fornia method (Eq. (5)) was applied for the calculation of RI of extreme
extreme drought, recurrence interval of severe drought, annual rainfall, and severe drought (Saha et al., 2021).
rainfall trend, critical rainfall, and wet-day frequency (Table 1). For the n
calculation of drought frequencies of extreme and severe drought, mean RI = (Eq. 5)
Er
drought intensity (MID), drought magnitude (MD), and recurrence in­
terval of extreme and severe drought Standard Precipitation Index (SPI) where, n denotes the number of events, Er is the order or rank of the
is used (Fig. 3).

4
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 3. The spatial data layers prepared for the indicators of Exposure: (a) Annual rainfall, (b) rainfall trend, (c) mean drought intensity, (d) drought magnitude, (e)
severe drought frequency, (f)extreme drought frequency, (g) return period of severe drought, (h) return period of extreme drought, (i) critical rainfall and (j) wet-
day frequency.

event.
Table 2
The rainfall trend was determined using linear regression. Critical
Classification of Drought according to SPI values.
rainfall or threshold rainfall can be calculated using Eq. (6) (Saha et al.,
Values Drought classes 2021):
More than 0 Non-Drought
0 to − 1.0 Mild Drought CR = σSPI + X (Eq. 6)
− 1.0 to − 1.5 Moderate Drought
− 1.5 to − 2.0 Severe Drought where, σ represent the Standard Deviation, SPI is equal to − 1.5, X ‾
Less than − 2 Extreme Drought defines the mean value. In this work the SPI value “-1.5” was selected as
critical rainfall values.

Table 3 2.3.2. Sensitivity indicators

SPI duration and their applications. The sensitivity indicators included in our present work are the
SPI Phenomena reflected Application average temperature, potential evapotranspiration (PET), net sown
duration area, percentage of small farmers, cultivable area, cropping intensity,
1 month Short term conditions Short term soil moisture and crop stress crop water demand, total water use, and population density (Fig. 4)
(especially during the growing season) (Table 4). These factors influence the responses to the potential threats
3 months Short- and medium-term A seasonal estimation of precipitation of being exposed. For instance, more population pressure in an area
moisture condition means more people are exposed to drought thereby increasing its
6 months Medium term trends of Potential for effectively showing the
vulnerability (Cooley et al., 2019). Also, the potential evapotranspira­
precipitation precipitation over distinct season
9 months Precipitation pattern over If SPI 9 < 1.5 then it is a good indication tion and temperature are the most valuable meteorological indicators
a medium time scale that substantial impact occurs in that have an impact on land cover, water balance, and ecological sus­
agriculture (and possibly other sectors) tainability (Chen et al., 2007; Kusangaya et al., 2014; Leta et al., 2016).
12 Long term precipitation Possibly tied to stream flows, reservoirs
This category of sensitivity includes both biophysical and economic
months pattern levels, and also ground water levels
factors.

2.3.3. Resilience indicators

Vulnerability is typically denoted as the inability, of a particular

5
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 4. The spatial data layers prepared for the indicators of Sensitivity: (a) crop water demand, (b) Cultivable land, (c) total water use, (d) population density, (e)
average temperature, (f) PET, (g) net area sown (hectares), (h) cropping intensity (%) and (i) small farmers (%).

population group, to adequately respond to a specific pernicious stressor membership type. The large fuzzy membership function is calculated
(Rygel et al., 2006). Therefore, the social vulnerability of a social group using Eq. (7) and the small membership function can be calculated using
is ascertained by the lack of resources through which they can respond Eq. (8).
to a hypothetical hazard (Johnson et al., 2012). Hence, the
1
socio-economic indicators of resilience i.e., forest area, water avail­ F(x) = ( ) (Eq. 7)
ability, net irrigated area, income index, education index, health index, 1+ f
x
2 − f 1

and fertilizer consumption are included in this work (Table 5). These are
the factors that represent the stress-absorbing ability of the population 1
of an area or in other words the adapting or recovering capacity of the F(x) = ( ) (Eq. 8)
population from drought (Fig. 5). For example, the forest cover will 1 + f x2 f 1

indicate the greenery of the area and thus the awareness of the people
and government towards drought shocks. where, f1 defines the spread parameter indicating the character and
shape of the transition zone and f2 denotes the midpoint values, which
2.4. Standardization of predisposing indicator also indicates the higher probability of becoming a member of the set.
After that, the machine learning models (i.e., ANN and Bagging)
After defining the set of relatable parameters spatial layers were were used to prepare the exposer, sensitivity, and resilience index maps.
created in the ArcGIS environment. The exposer index, sensitivity index, Finally, the relative drought vulnerability map was developed using eq.
and resilience index were developed employing spatial data layers of (9) (Singha et al., 2020)
exposer indicators, sensitivity indicators, and resilience respectively.
Before initiating indexing, the prepared spatial layers are standard­ RDV = (E + S) − R (Eq. 9)
ized using the Fuzzy Membership tool in ArcGIS. Fuzzy membership is a
function that classifies the dataset on a scale of 0–1 and was introduced where, RDV define as relative drought vulnerability, R denotes the
by Zadeh (1965). In this function, out of Gaussian, Near, Linear, Small, resilience index, S defines the sensitivity index, and E denotes as an
Large, MSLarge, and MSSmall, the large and small functions were used exposure index. For measuring the role of sensitivity, exposure, and
in our present work as they efficiently define our sets of indicators. The 3 resilience all the parameters were normalized using Eq. (10) (for posi­
sets of layers were standardized according to their motive of sub-indices. tively correlated) and Eq. (11) (for negatively correlated).
For example, the layers which are directly related to resilience, sensi­ Xi − Xmin
Zi = (Eq. 10)
tivity, and exposure were assigned large in membership type whereas Xmax − Xmin
the layers which are indirectly related were assigned small as

6
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Table 4 Table 5
Showing the directionality of influence of indicators of sensitivity to drought Showing the directionality of influence of indicators of resilience to drought
vulnerability. vulnerability.
Sl. Variables Logical Directionality Source Sl. Variables Logical Directionality of Source
no. consideration of influence to no. consideration influence to
drought drought

1 Crop water demand More the crop Direct Statistical 1 Forest area More the forest is Indirect Statistical
water demand Handbooks less will be the Handbooks
more will be the vulnerability
vulnerability 2 Cultivable More the Cultivable Indirect Statistical
2 Water availability More the water Indirect Statistical area area less will be the Handbooks
availability less Handbooks vulnerability
will be the 3 Net irrigated More the net Indirect Statistical
vulnerability area irrigated area less Handbooks
3 Total water use More the total Indirect Statistical will be the
water useless Handbooks vulnerability
will be the 4 Health index More the value of Indirect HDI Report,
vulnerability health index less will 2012
4 Population density More the Direct Census of be the vulnerability
population India, 2011 5 Education More the value of Indirect HDI Report,
density more index education index less 2012
will be the will be the
vulnerability vulnerability
5 Average More will be the Direct IMD 6 Income index More the value of Indirect HDI Report,
Temperature temperature income index less 2012
more will be the will be the
vulnerability vulnerability
6 Potential More will be the Direct IMD 7 Fertilizer More the fertilizer Indirect Statistical
Evapotranspiration PET more will be consumption consumption less Handbooks
(PET) the vulnerability will be the
7 Net sown area More the net Indirect Statistical vulnerability
sown area less Handbooks
will be the
vulnerability the researchers have used the feed-forward based multilayer perceptron
8 Cropping intensity More the Indirect Statistical
(MLP) algorithm. The MLP was built by output, more hidden, and input
cropping Handbooks
intensity less
layers of non-linear activation nodes. The individual layer comprises of
will be the numerous neurons; those are attached with a certain weight to each
vulnerability node in the higher layer. Their results were transferred to get the in­
9 % of small farmers More the Direct Statistical formation. In this way, the ANN has been developed. The back­
percentage of Handbooks
propagation algorithm was applied in MLP for the construction of the
small farmers
more will be the network till the least error is reached within the output and anticipated
vulnerability values of the ANN.

2.5.2. Bagging
( )
Xi − Xmin Breiman introduced the Bagging algorithm in 1996. It is a well-
Zi = 1 − (Eq. 11)
Xmax − Xmin known machine learning algorithm that uses a bootstrap sampling
technique (Breiman, 1996). Several researchers have applied this model
Where, Xi indicates the specific value of a parameter, Xmax and Xmin to predict susceptibility assessment (i.e., landslide, deforestation, etc.)
represent the maximum and minimum values of the X parameter. because this model has a good prediction capacity (Wu et al., 2020; Saha
et al., 2021a; Hong et al., 2020). For the assessment of drought
vulnerability, several models are applied, but this algorithm has not yet
2.5. Machine learning models for the RDV been applied by researchers. This algorithm is selected because it fab­
ricates the decision tree with the help of each produced subsets and
2.5.1. Artificial neural network (ANN) finally, they are assembled within the final model (Hong et al., 2020). It
The ANN is a black box and a mathematical model. These ANN is enhancing the alignment accuracy by minimizing the inconsistency of
models can solve the clustering problems which are generated during the alignment error (Saha et al., 2021a; Wu et al., 2020). Bagging
the implementation of Bayesian decision theory. These networks were classifier is considered as three steps bagging system (Breiman, 1996;
developed by the implementation of two layers. In this implemented Saha et al., 2021a; Yariyan et al., 2020). It is developed as a bootstrap
network the second layer, dissimilar to other networks, targets the sample by the substantive training samples through the displacement
outcomes in the vector form with values ranging between 1 and 0 (Azimi approach (Saha et al., 2021a). This machine learning model can promote
et al., 2018). Therefore, it can work like a human brain and can robustly the success of all array of subsets by connecting them to the actual
predict the result from a large number of variables. This is the reason feature process for the bagging classification stage (Breiman, 1996;
why, number of researchers have extensively applied this algorithm in Yariyan et al., 2020).
several fields of research (i.e., landslide, forest fire, flood, gully erosion
vulnerability assessment, etc.). The ANN algorithm can apply to the
binary, category, and continuous without violating the characters and 2.6. Validation of RDV models
assumptions of the data. It can act as an expert, which has the predictive
capacity to detect the complicated predictive pattern, that is non-expert 2.6.1. Receiver operating characteristics curve (ROC)
not apparent (Wang et al., 2016). Many researchers have widely applied Validation is an important task to evaluate the scientific significance
this model (Termeh et al., 2018; Huang et al., 2018; Harmouzi et al., of an implemented study (Gayen and Saha, 2017; Talukdar and Pal,
2019; Moayedi et al., 2019; Falah et al., 2019; Zhao et al., 2019). Here, 2020). Here, the researchers have applied the AUC-ROC to validate the

7
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 5. The spatial data layers prepared for indicators of Resilience: a. forested area, b. water availability, c. net irrigated area, d. health index, e education index, f.
income index, g. fertilizer consumption.

prediction capacity of the algorithms. The ROC is a pictorial illustration respectively (Abedinpour et al., 2012; Garosi et al., 2019). Efficiency is
of model performances through a diagnostic test (Singha et al., 2020; the ratio between drought non-occurrence and drought occurrence
Saha et al., 2021). The ROC shows the model sensitivity on the Y-axis by pixels that are accurately classified. It was predicted by applying Eq.
plotting the true predictions (drought-affected location) and the false (12) (Rahmati et al., 2017).
predictions (non-drought effected location) shown on the X-axis (Roy
TP + TN
and Saha, 2020; Saha et al., 2020a). The models’ prediction was eval­ Efficiency = (Eq. 12)
TP + TN + FP + FN
uated by the AUC. The value of AUC is extended from 0.5 to 1; the value
closer to 1 indicates a good prediction capacity of models (Rasyid et al.,
2.6.3. Reliability accuracy measures by RMSE and MAE
2016; Gayen et al., 2020).
The efficiency was applied to identify the prediction power of the
implemented models while the RMSE and MAE assess the error in pre­
2.6.2. Efficiency
dictive models (Abedinpour et al., 2012). The RMSE was calculated by
The efficiency (E) is applied as the error tests to assess the validation
the comparison between field observed values and model generated
and calibration results of the applied models (Nash and Sutcliffe, 1970;
predicted values. The RMSE value was calculated using Eq. (13) (Rah­
Abedinpour et al., 2012). Employing various cut-off data, the predicted
mati et al., 2017; Saha et al., 2020):
results were subdivided into one of the explanation levels (i.e., drought
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
absence and presence). For defences cut-off data, there must be present 1
some wrong classification status (Garosi et al., 2019). So, there are true RMSE = ∑ (Eq. 13)
(N) Ni=1 (O i − S i )2
negative (TN) and true positive (TP) classes are found, which means
some pixels are recognized as drought occurrences and non-occurrence, MAE is the same as RMSE, it is calculated as the aggregate of dif­
respectively. There is also the presence of false negative (FN) and false ferences between model predicted values and field observed values,
positive (FP), which also represents the number of drought pixels that excepting their direction (Eq. (14)).
were wrongly classified as non-occurrence of drought and vice versa,

8
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ n
(S i − O i ) 3.2. RDV modelling using ANN and bagging models
MAE = (Eq. 14)
n
i=1
The RDV maps were developed using the two well-accepted MLAs
(Fig. 8a and b). Models of drought vulnerability utilizing ML frameworks
where Oi and Si are observed and predicted values. N denotes the total
were designed using training data sets to minimize drought vulnerability
number of observations.
in the area being studied. Continuous variables were employed to
perform the ML models. The training datasets for drought and non-
2.6.4. Assessment of true skill statistics (TSS)
drought were considered. The randomly set points (70%) from both
The Hanssen-Kuipers discriminate, also named True skill statistics, is
classes were used to train models. After two models were properly
widely accepted in the field of weather forecasting to assess the level of
implemented in the training phase, the relational weights of the models
prediction accuracy (Accadia et al., 2005; Pourghasemi et al., 2020). In
were used to calculate the drought vulnerability indices for the whole
the field of the confusion matrix, it is formulated as Eq. (15):
field of study. By using the trial-and-error approach to develop such
TSS = Sei + Spi − 1 (Eq. 15) drought vulnerability models, the measurement variables were stan­
dardized. In this work, ArcGIS and R-studio were applied to evaluate the
where, Spi and Sei are defined as specificity and sensitivity, respectively drought vulnerability using ANN and bagging models. For this test, 1
(Pourghasemi et al., 2020). The value of TSS was extended from 1 to − 1, hidden layer, 0.3 learning rate, 0.2 dynamics, 0 seeds, 500 training times
with values nearest to 1 that’s means excellent model discrimination and 20 thresholds of validation were utilized for ANN. For bagging, 16
was found within background and occurrence locations and <0.4 rep­ iterations, 1 seed, 100% of bag size (training range size), and a reduced
resenting poor model discrimination (Pourghasemi et al., 2020). error pruning tree were used as basic classifiers. The relative drought
vulnerability as predicted by each model is categorized into five
2.6.5. Friedman and Wilcoxon rank test vulnerability classes with the help of natural-break statistics (Gayen
Friedman developed a non-parametric test in 1937, for the assess­ et al., 2019; Pourghasemi et al., 2020). The RDV developed with the
ment of significant differences within two implemented models (Singha ANN model exhibits values ranging from 1.0 to 0.0. The high and very
et al., 2020). If the data will normally be distributed, we can apply this high drought vulnerability zones cover 9.74% and 52.03% of the study
statistic (Martinez-Alvarez et al., 2013). Before going to start the area, respectively. Very low, low, and moderate vulnerability zones
Friedman test, all the predictive models should be at a 5% significance make up 16.64%, 13.09%, and 8.50% of the total area (Table 6). This
level (p < 0.05). In case the P-value of both the models are defined as Bagging vulnerability result reveals that 40.87% and 16.69% of the
true in the Friedman test, that means there is no significant difference Karnataka state fall under very high and high zones for drought occur­
within their prediction (Bui et al., 2015; Singha et al., 2020). The Wil­ rences. The very low, low, and moderate RDV classes cover 15.35%,
coxon rank test was applied to solve this problem by which the re­ 17.03%, and 10.06% of the Karnataka state in India, respectively
searchers can check the methodical level of significant differences (Table 6). It is clear in all drought vulnerability maps; the
within the RDV models (Bui et al., 2015). If the z-value was greater than drought-affected areas are found in the Central and Eastern parts of this
+1.96 or − 1.96 with p < 0.05, the alternative hypothesis was accepted; state and the least vulnerable level is found in the southwestern part of
so, the models were significantly different (Singha et al., 2020). Karnataka. Concurrent with the output from both the models, it is shown
that ANN and Bagging models generate RDV maps that are smooth
3. Results and analysis patterns. Both models reveal that Gulbarga, Tumkur, Bijapur, Bellary,
Chitradurg, Raichur, and Bagalkote districts are more vulnerable to
3.1. Analysing the contribution of indicators drought occurrences. Besides, the less vulnerable districts are Mysore,
Shimoga, Bangalore, and Belgaum. Biradar and Sridhar (2009) reported
It is a vital task to build the relationships between the predisposing that more than 80% of total people were impacted by drought impact
factors and drought-affected locations to assess the impact of each driver during 2002–03. Those people were forced to leave agricultural activ­
in the intensification of drought vulnerability (Fig. 6). This map gives us ities to join the labour force outside villages and in the relief sites. They
adequate information for a comparison of drought factors between the used to live in Gulbarga, Tumkur, Bellary, and Bagalkote districts. These
Karnataka districts. Generally, drought vulnerability is impacted by districts fall under the most vulnerable areas as per the prediction result
different predisposing factors (Chen et al., 2007; Yuan et al., 2015). The thereby attesting to the model’s prediction capacity.
normalized values range from 0 to 1. Also, bar graphs were prepared for
every district of Karnataka, to represents the indicators and their relative 3.3. Spatial association of resilience, sensitivity, and exposure models
contribution to drought vulnerability. The orange, red, and green col­ with RDV models
ours correspond to resilience, sensitivity, and exposure markers,
respectively. The result reveals that the indicators of sensitivity domi­ The resilience, sensitivity, and exposure models are correlated with
nantly contributed to drought vulnerability in Bagalkote district, but in RDV models for the assessment of their impact on vulnerability occur­
the case of Yadgir district the exposure indicators play an important role rences (Singha et al., 2020). From the exposure map, it is found that the
in drought vulnerability. In most cases, the water availability, total western part of Karnataka is least exposed to drought vulnerability
water use, drought magnitude of 12 months, the return period of severe whereas the North, Central, and Southern part of Karnataka are highly
drought, forest area, and income index were the most important factors. exposed to vulnerable conditions. The sensitivity map of drought
These were followed by annual rainfall, net irrigated area, the return vulnerability conforms to the results of the exposure map - that western
period of extreme drought, and average temperature as influencing Karnataka is least sensitive to drought vulnerability whereas
vectors. All the 26 predisposing factors were subjected to modelling south-eastern Karnataka is most sensitive. The resilience map also shows
purpose because all are contributors to drought vulnerability occur­ that the north-western and the southern part of Karnataka are efficiently
rence. The maximum area of the eastern part of this state falls under less capable of adaptation whereas the north-eastern part of Karnataka is the
water availability, low critical rainfall with low income and education least capable (Fig. 7a–c). The overall RDV and resilience-based RDV had
index, and high PET value that is favourable for drought vulnerability an important role in RDV because there was a significant spatial asso­
occurrences. ciation between them. The sensitivity, exposure, and resilience index
were correlated with RDV results to assess their impacts. Here, Karl
Pearson’s (1897) correlation method was implemented for the assess­
ment of the degree and direction of the relationship between

9
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 6. Bar graph showing the contribution of indicators under each category of exposure, sensitivity and resilience.
Note: 1 = Annual rainfall, 2 = Rainfall trend, 3 = Wet day frequency, 4 = Critical rainfall, 5 = Mean drought intensity of 12 months, 6 = Drought magnitude of 12 months,
7 = Extreme drought frequency of 12 months, 8 = Severe drought frequency of 12 months, 9 = Return period of extreme drought of 12 months, 10 = Return period of severe
drought of 12 months, 11 = Average temperature, 12 = Potential evapotranspiration, 13 = Net sown area, 14 = percentage of small farmers, 15 = Water availability, 16 =
Cropping intensity, 17 = Crop water demand, 18 = Total water use, 19 = Population density, 20 = Forest area, 21 = Cultivable area, 22 = Net irrigated area, 23 = Income
index, 24 = Education index, 25 = Health index, 26 = Fertilizer consumption.

10
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Table 6 linearly-related variables. The correlation values extend from − 1 to +1,

Areal share under vulnerability classes of RDV, exposure, sensitivity, and resil­ where − 1 and +1 indicate strongly negative and positive relationships
ience models (area in %). (Pourghasemi et al., 2020). The results revealed that all the indices are
Indices Model Very high High Moderate Low Very low statistically significant (Table 7). In all the cases the correlation values
Exposure ANN 50.57 8.54 5.67 9.09 26.13
vary from 0.612 to 0.995. Exposure index is highly correlated with RDV
Bagging 46.07 12.13 10.50 8.56 22.74 model (r = 0.955–0.958), followed by the sensitivity (r = 0.895–0.944),
Sensitivity ANN 56.80 5.34 4.99 6.31 26.55 and resilience (r = 0.612–0.672).
Bagging 48.45 10.69 10.31 7.55 22.99
resilience ANN 54.76 6.68 5.12 8.12 25.31
3.4. Validation and comparisons of drought vulnerability models
Bagging 37.61 15.86 13.38 12.23 20.92
RDV ANN 52.03 9.74 8.50 13.09 16.64
Bagging 40.87 16.69 10.06 17.03 15.35 Accuracy assessment is an important task to build a conclusion about
the prediction capacity of the implemented models (Saha et al., 2020b).
For comparing and validating the implemented models the ROC, RMSE,
MAE, efficiency, Friedman test, and Wilcoxon test were implemented.
The results of the ROC curve indicate that the ANN and Bagging models
had 84.6% and 82.9% accuracy in cases of success rate curve and 86.5%
and 84.2% accuracy in cases of prediction rate curve, respectively; the
significance level is 0.00 (Fig. 9). The ROC values range from 0.83 to
0.86, demonstrating that both models have excellent drought vulnera­
bility prediction capacity. TSS, RMSE, MAE and efficiency values of the
relative drought vulnerability models have justified the good capability
of the applied MLAs (Table 8). The sensitivity, exposure, and resilience
indices were correlated with RDV results to assess their impacts. The
results of Friedman, and Wilcoxon test results of the resilience, sensi­
tivity, and exposure are presented in Tables 9 and 10 for both the
models.
The Friedman and Wilcoxon tests revealed that the Bagging model
computed the greatest mean rank of overall RDV, despite resilience,
sensitivity, and exposure indices (Table ss 9 and 10). The calculated Chi-
square value ranges from 11.76 to 1 at the 3rd degree of freedom also
shows that there were slight differences between the models, so, all the
models can be expressed at a uniform satisfactory level. Also, we can
observe that the RMSE, TSS, and MAE values are nearly equal for both
the cases of ANN and Bagging models. Thus, we can conclude that both
models have uniform prediction capacity for the measurement of
drought vulnerability.

4. Discussion

For the drought vulnerability assessment factors like extreme

drought frequency (%), severe drought frequency (%), recurrence in­
terval of extreme and severe drought, mean drought intensity, drought
magnitude, critical rainfall, annual rainfall, rainfall trend, potential
evapotranspiration (PET), average temperature, wet-day frequency, net
sown area, net irrigated area, population density, percentage of small
farmers, cropping intensity, water availability, crop water demand, total
water use, forest area, cultivable area, income index, education index,
the health index, and fertilizer consumption is employed. In similar
works (Ghosh, 2019; Jiang et al., 2019; Mundetia and Sharma, 2015;
Chandrasekar et al., 2009; Sridevi et al., 2014; Thomas et al., 2016;
Lindner et al., 2010; Udmale et al., 2014; Ortega-Gaucin et al., 2018; Ebi
and Bowen, 2016; Siebert et al., 2019) the same set of factors are used
for evaluating drought vulnerability and the relation between the factor
used and the drought is also found to be same. Several researchers have
developed drought vulnerability maps for strategic adaption, drought
risk reduction, and long-term planning purposes (Ahmadalipour et al.,
2018; Hagenlocher et al., 2019; Hoque et al., 2020).
Several researchers also applied MLAs like ANN and Bagging, in
various disciplines, like a landslide, flood hazard, gully erosion, and
Fig. 7. Exposure index, Sensitivity index, and Resilience index maps produced deforestation susceptibility assessment (Falah et al., 2019; Arabameri
by ANN and Bagging models. et al., 2020; Ghorbanzadeh et al., 2019; Hong et al., 2020; Roy and Saha,
2020; Saha et al., 2020; Talukdar et al., 2020; Yu and Chen, 2020). In
every case, the prediction capacity of the model’s result is extremely
appreciated. Thus, the application of machine learning algorithms
(MLAs) is not a new thing, but the implication of these MLAs for relative
drought vulnerability assessment is rather unique. Previous research

11
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Fig. 8. The Relative Drought Vulnerability maps produced by a. ANN and b. Bagging models.

function approximation problem may be solved with ANN (Park and

Table 7
Sandberg, 1993; Yilmaz et al., 2011). Bagging model has also good ac­
Co-relation of exposure, sensitivity, and resilience of ANN and Bagging models.
curacy in predicting the drought vulnerability. The Bagging fabricates
Drought indicators Relative drought vulnerability the decision tree with the help of each produced subsets, and ultimately
ANN Bagging assembles them within the final model (Hong et al., 2020). The potential
Exposure 0.958 0.955 benefits of applied MLAs, according to Yilmaz et al. (2011), go beyond
Sensitivity 0.944 0.895 fast computation rates. Because there are many more processing neu­
Resilience 0.612 0.672 rons, each with largely local connections, these models perform better
than conventional statistical models in terms of reliability and accuracy.
The performance comparison also revealed that applied MLAs are
work likes Gayen et al. (2019); Saha et al. (2020); Singha et al. (2020);
effective instruments for reducing uncertainties, and their application
Pal et al. (2020) concluded that MLAs provided adequate results con­
might lead to new ideas and methodologies as well as a reduction in
cerning binary statistical models. In other studies, like floods, landslides,
correlation inconsistencies.
and assessments of spring potential, the ANN model has also given good
Complete reduction of overfitting problem is not possible for many
results. In the case of no overfitting the data, the ANN model is the most
causes. First, the testing and training data sets had a 70:30 sampling
important.
ratio with no sampling accuracy checks. Second, the drought vulnera­
The ANN has a slightly higher prediction accuracy of 2.3% than the
bility indicators employed in this study may contain noise. Finally, if the
bagging model, according to the validation findings. The feed-forward
number of chosen drought vulnerability indicators changes, drought
artificial neural network is the most often used algorithm. Any

Fig. 9. Validation of Relative Drought Vulnerability maps applying ROC curve: (a) Success rate curve (applying training dataset) and (b) Prediction rate curve
(applying validation dataset).

12
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Table 8 sector. To reduce the loss of resources caused by the drought of the study
Estimation of efficiency, area under the curve, true skill statics, root mean square area first and fore more work is to identify the area vulnerable to
error, and MAE for both machine learning algorithms. drought and proper drought controlling strategies are also essential. In
Matrices Training dataset Validation dataset the present study drought vulnerability mapping has been carried out
ANN Bagging ANN Bagging
using advanced machine learning that is new in this field and could be
used as an important tool to predict the vulnerable area to drought. The
Efficiency 0.85 0.83 0.89 0.88
present work would help the government and non-government organi­
AUC 0.846 0.829 0.865 0.842
TSS 0.79 0.78 0.82 0.81 zation in implementing suitable drought management strategies such as
RMSE 0.29 0.33 0.32 0.33 rainwater harvesting, deployment of irrigational infrastructures, affor­
MAE 0.39 0.43 0.46 0.48 estation and reducing the effect of drought in the specific drought
vulnerable area.

Table 9 5. Conclusion
Mean rank values of the computed models in Friedman test.
Indices Mean rank Chi square P value In the present study of drought vulnerability assessment, a definite
methodology is used and applied in Karnataka which is yet not used in
ANN Bagging
the drought vulnerability assessment. Since the character of drought
Exposure 1.39 1.61 6.627 0.007 varies the method for assessing drought vulnerability concomitantly
Sensitivity 1.49 1.51 6.561 0.009
varies across space and time. Based on previous studies, twenty-six
Resilience 1.5 1.5 1.00 1.00
RDV 1.36 1.64 11.76 0.001 drought effective factors were overlaid on a GIS platform for RDVMs.
These factor layers were integrated with the ANN and Bagging models.
According the validation statistics ANN model produced better result
than the Bagging method. The ANN and Bagging models identified
Table 10
52.03% and 40.87%, respectively, of the study area to have a very high
Wilcoxon test for pair wise comparison of exposure, sensitivity, resilience, and
relative drought vulnerability models. drought vulnerability. The results indicate that this region would
certainly face a vulnerable situation in the future if proper drought
Indices Z value P value Significance
management schemes are not implemented. The study is limited in
ANN vs Bagging terms of non-operationalizing of field observations and recording local
Exposure − 4.754 0 Yes people’s perceptions for lack of funding and time constraints. Another
Sensitivity − 2.79 0.005 Yes limitation of the work is the absence of agriculture dependency data.
Resilience − 0.882 0.378 No However, such a shortcoming doesn’t impact the accuracy level of the
RDV 4.054 0 Yes
applied models. Finally, this research approach has an excellent scope
−

for the delineation of the RDV zones by switching the associated datasets
vulnerability models may alter. Another flaw in the models is that the and variables for further planning and development in other areas. From
drought vulnerability indicators assumed do not change over time. This the above study, it may be concluded that Karnataka receives heavy
is a typical flaw in many machine learning models, but it’s particularly southwest monsoon rainfall but it is only limited to the coastal areas.
relevant to ours because drought vulnerability indicators were picked Karnataka being in the Deccan Trap consists of huge masses of igneous
based on predisposing risk factors for drought vulnerability (Geist and geology due to which the chances of infiltration are low and the wide­
Lambin, 2001; Mas et al., 2004). The machine learning methods spread urbanisation and concretisation further accentuate the condition.
employed in this work have the benefit of making the process of ana­ Since natural infiltration is difficult people there should practice artifi­
lysing various databases in order to acquire essential data much easier. cial rainwater harvesting. The government should introduce various
Such algorithms may accommodate for certain assumptions and can be schemes and generate awareness among common people. Karnataka
used in conjunction with automated data analysis to aid should install various types of machinery and various irrigational in­
decision-making. The following could be included in future studies to frastructures which will help them in conserving water as they are
improve the accuracy of drought vulnerability models like analysis of already very advanced in technology. The finding of this research can be
the ratio of training and validation datasets, application of various used for water resource management, environmental protection, and
feature selection strategies, more effective drought vulnerability in­ land use planning by private and local governmental agencies in the
dicators and model output review using specific approaches. The dif­ Karnataka state of India.
ference in exposure to vulnerability may be the effect of the presence of
Western Ghats. Because of the presence of Western Ghats, the coastal Funding
areas receive heavy downpour whereas the immediate rain shadow
zones remain dry. The south-eastern part of Karnataka comprises India’s No fund was received for this work.
5th most populated urban agglomeration (Sudhira et al., 2007). There­
fore a large population resides here, thus making it the most sensitive. CRediT authorship contribution statement
The widespread concretisation and urbanisation has resulted in acute
water shortages thereby increasing the sensitivity. The capability to Sunil Saha: Methodology, Formal analysis, Writing – original draft,
adapt also varies due to the difference in forest cover. Finally, it could be Writing – review & editing. Priyanka Gogoi: Methodology, Formal
concluded that the overall western Karnataka falls under least exposed analysis, Investigation, Software, Writing – original draft. Amiya
and least sensitive but highly adaptive to drought whereas eastern Gayen: Methodology, Formal analysis, Writing – original draft, Writing
Karnataka is highly exposed and sensitive and least adaptive to drought. – review & editing. Gopal Chandra Paul: Methodology, Formal anal­
Thus, it could be concluded that the presence of Western Ghats has a ysis, Investigation, Software, Writing – original draft.
great impact on welcoming the monsoons and influencing the resultant
drought conditions. Frequent occurrence of drought is an important Declaration of competing interest
barrier for the environmental development as well as socio-economic
condition where most of the people of India rely on the agriculture The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence

13
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

the work reported in this paper. Huang, B., Zhao, B., Song, Y., 2018. Urban land-use mapping using a deep convolutional
neural network with high spatial resolution multispectral remote sensing imagery.
Remote Sens. Environ. 214, 73–86.
References IPCC, 2001. Summary for Policymakers. Climate Change 2001: Impacts, Adaptation, and
Vulnerability: A Report of Working Group II of the IPCC. IPCC, Geneva.
Abedinpour, M., Sarangi, A., Rajput, T.B.S., Singh, M., Pathak, H., Ahmad, T., 2012. IPCC, 2007. Climate change 2007: impacts, adaptation and vulnerability. In: Parry, M.L.,
Performance evaluation of AquaCrop model for maize crop in a semi-arid Canziani, O.F., Palutikof, J.P., van der Linder, P.J., Hanson, C.E. (Eds.), Contribution
environment. Agric. Water Manag. 110, 55–66. of Working Group II to the Fourth Assessment Report of the IPCC. Cambridge
Accadia, C., Mariani, S., Casaioli, M., Lavaqnini, A., Speranza, A., 2005. Verification of University Press, Cambridge.
precipitation forecasts from two limited-area models over Italy and comparison with Jafari Shalamzari, M., Zhang, W., Gholami, A., Zhang, Z., 2019. Runoff harvesting site
ECMWF forecasts using a resampling technique. Weather Forecast. 20, 276–300. suitability analysis for wildlife in Sub-Desert regions. Water 11 (9), 1944.
Agnew, C.T., 1990. Spatial aspects of drought in Sahel. J. Arid Environ. 18, 279–293. Jiang, W., Wang, L., Feng, L., Zhang, M., Yao, R., 2019. Drought characteristics and its
Arabameri, A., Asadi Nalivan, O., Chandra Pal, S., Chakrabortty, R., Saha, A., Lee, S., impact on changes in surface vegetation from 1981 to 2015 in the Yangtze River
Tien Bui, D., 2020. Novel machine learning approaches for modelling the gully Basin, China. Int. J. Climatol. 40 (7), 3380–3397.
erosion susceptibility. Rem. Sens. 12 (17), 2833. Johnson, D.P., Stanforth, A., Lulla, V., Luber, G., 2012. Developing an applied extreme
Azimi, S.M., Britz, D., Engstler, M., Fritz, M., Mücklich, F., 2018. Advanced steel heat vulnerability index utilizing socioeconomic and environmental data. Appl.
microstructural classification by deep learning methods. Sci. Rep. 8 (1), 1–14. Geogr. 35 (1–2), 23–31.
Biradar, N., Sridhar, K., 2009. Consequences of 2003 drought in Karnataka with Kelly, P.M., Adger, W.N., 2000. Theory and practice in assessing vulnerability to climate
particular reference to livestock and fodder. J. Hum. Ecol. 26 (2), 123–130. change andFacilitating adaptation. Climatic Change 47 (4), 325–352.
Breiman, L., 1996. Bagging predictors. Mach. Learn. 24, 123–140. Kusangaya, S., Warburton, M.L., Van Garderen, E.A., Jewitt, G.P., 2014. Impacts of
Bui, T.T., Tu, H.P., Dang, M.C., 2015. DRIE process optimization to fabricate vertical climate change on water resources in southern Africa: a review. Phys. Chem. Earth,
silicon nanowires using gold nanoparticles as masks. Adv. Nat. Sci. Nanosci. Parts A/B/C 67, 47–54.
Nanotechnol. 6 (4), 045016. Lakshmi, S.V., Ramalakshmi, M., Rakshith, R.K., Christobel, M.J., Kumar, P.P.,
Chandrasekar, K., Sai, M.S., Roy, P.S., Jayaraman, V., 2009. Identification of agricultural Priyadharshini, B., Kumar, P.R., 2020. An integration of geospatial technology and
drought vulnerable areas of Tamil Nadu, India–Using GIS based multi criteria standard precipitation index (SPI) for drought vulnerability assessment for a part of
analysis. Asian Journal of Environment and Disaster Management 1 (1). Namakkal district, South India. Mater. Today: Proceedings 33, 1206–1211.
Chen, H., Guo, S., Xu, C.Y., Singh, V.P., 2007. Historical temporal trends of hydro- Leta, O.T., El-Kadi, A.I., Dulai, H., Ghazal, K.A., 2016. Assessment of climate change
climatic variables and runoff response to climate variability and their relevance in impacts on water balance components of Heeia watershed in Hawaii. J. Hydrol.: Reg.
water resource management in the Hanjiang basin. J. Hydrol. 344 (3–4), 171–184. Stud. 8, 182–197.
Cheng, J., Tao, J.P., 2010. Fuzzy comprehensive evaluation of drought vulnerability Lin, H., Wang, J., Li, F., Xie, Y., Jiang, C., Sun, L., 2020. Drought trends and the extreme
based on the analytic hierarchy process:—an empirical study from Xiaogan City in drought frequency and characteristics under climate change based on SPI and HI in
Hubei Province. Agriculture and Agricultural Science Procedia 1, 126–135. the upper and middle reaches of the huai river basin, China. Water 12 (4), 1100.
Cooley, S.S., Williams, C.A., Fisher, J.B., Halverson, G.H., Perret, J., Lee, C.M., 2019. Lindner, M., Maroschek, M., Netherer, S., Kremer, A., Barbati, A., Garcia-Gonzalo, J.,
Assessing regional drought impacts on vegetation and evapotranspiration: a case Lexer, M.J., 2010. Climate change impacts, adaptive capacity, and vulnerability of
study in Guanacaste, Costa Rica. Ecol. Appl. 29 (2) e01834. European forest ecosystems. For. Ecol. Manag. 259 (4), 698–709.
Dracup, J.A., Lee, K.S., Paulson Jr., E.G., 1980. On the definition of droughts. Water Liu, C., Yang, C., Yang, Q., Wang, J., 2021. Spatiotemporal drought analysis by the
Resour. Res. 16 (2), 297–302. standardized precipitation index (SPI) and standardized precipitation
Ebi, K.L., Bowen, K., 2016. Extreme events as sources of health vulnerability: drought as evapotranspiration index (SPEI) in Sichuan Province, China. Sci. Rep. 11 (1), 1–14.
an example. Weather and Climate Extremes 11, 95–102. Liu, X., Wang, Y., Peng, J., Braimoh, A.K., Yin, H., 2013. Assessing vulnerability to
Edwards, D.C., 1997. Characteristics of 20th Century Drought in the United States at drought based on exposure, sensitivity and adaptive capacity: a case study in middle
Multiple Time Scales (No. AFIT-97–051). Air Force Inst. of Tech Wright-Patterson Inner Mongolia of China. Chin. Geogr. Sci. 23 (1), 13–25.
AFB OH. Mandal, I., Pal, S., 2020. Modelling human health vulnerability using different machine
Engström, J., Jafarzadegan, K., Moradkhani, H., 2020. Drought vulnerability in the learning algorithms in stone quarrying and crushing areas of Dwarka river Basin,
United States: an integrated assessment. Water 12 (7), 2033. Eastern India. Adv. Space Res. 66 (6), 1351–1371.
Falah, F., Rahmati, O., Rostami, M., Ahmadisharaf, E., Daliakopoulos, I.N., Martinez-Alvarez, F., Reyes, J., Morales-Esteban, A., Rubio-Escudero, C., 2013.
Pourghasemi, H.R., 2019. Artificial neural networks for flood susceptibility mapping Determining the best set of seismicity indicators to predict earthquakes. Two case
in data-scarce urban areas. In: Spatial Modeling in GIS and R for Earth and studies: Chile and the Iberian Peninsula. Knowl.-Based Syst. 50, 198–210.
Environmental Sciences. Elsevier, pp. 323–336. Mas, J.F., Puig, H., Palacio, J.L., Sosa-Lopez, A., 2004. Modelling deforestation using GIS
Fraser, E.D.G., 2007. Travelling in antique lands: using past famines to develop an and artificial neural networks. Environ. Model. Software 19 (5), 461–471.
adaptability/resilience framework to identify food systems vulnerable to climate McKee, T.B., Doesken, N.J., Kleist, J., 1993. January. The relationship of drought
change. Climatic Change 83 (4), 495e514. frequency and duration to time scales. Proceedings of the 8th Conference on Applied
Garosi, Y., Sheklabadi, M., Conoscenti, C., Pourghasemi, H.R., Van Oost, K., 2020. Climatology 17 (22), 179–183.
Assessing the performance of GIS-based machine learning models with different Moayedi, H., Mehrabi, M., Mosallanezhad, M., Rashid, A.S.A., Pradhan, B., 2019.
accuracy measures for determining susceptibility to gully erosion. Sci. Total Environ. Modification of landslide susceptibility mapping using optimized PSO-ANN
644, 1117–1132. technique. Eng. Comput. 35 (3), 967–984.
Gayen, A., Haque, S.M., Saha, S., 2020. Modeling of gully erosion based on random forest Mundetia, N., Sharma, D., 2015. Analysis of rainfall and drought in Rajasthan State,
using GIS and R. In: Gully Erosion Studies from India and Surrounding Regions. India. Global Nest J 17 (1), 12–21.
Springer, Cham, pp. 35–44. Nageswara Rao, P.P., Shobha, S.V., Ramesh, K.S., Somashekhar, R.K., 2005. Satellite-
Gayen, A., Pourghasemi, H.R., Saha, S., Keesstra, S., Bai, S., 2019. Gully erosion based assessment of agricultural drought in Karnataka state. Journal of the Indian
susceptibility assessment and management of hazard-prone areas in India using society of remote sensing 33 (3), 429–434.
different machine learning algorithms. Sci. Total Environ. 668, 124–138. Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models. I. A
Gayen, A., Saha, S., 2017. Application of weights-of-evidence (WoE) and evidential belief discussion of principles. J. Hydrol. 10, 282–290.
function (EBF) models for the delineation of soil erosion vulnerable zones: a study on Ortega-Gaucin, D., De la Cruz Bartolón, J., Castellano Bahena, H.V., 2018. Drought
Pathro river basin, Jharkhand, India. Modeling Earth Systems and Environment 3 vulnerability indices in Mexico. Water 10 (11), 1671.
(3), 1123–1139. Pai, D.S., Bhan, S.C., 2015. Monsoon 2014: A Report. IMD Met. Monograph: ESSO
Geist, H.J., Lambin, E.F., 2001. What drives tropical deforestation. LUCC Rep. Ser. 4, Document No.: ESSO/IMD/SYNOPTIC MET/01 (2015)/17.
116. Pal, S., Kundu, S., Mahato, S., 2020. Groundwater potential zones for sustainable
Ghorbanzadeh, O., Valizadeh Kamran, K., Blaschke, T., Aryal, J., Naboureh, A., Einali, J., management plans in a river basin of India and Bangladesh. J. Clean. Prod. 257,
Bian, J., 2019. Spatial prediction of wildfire susceptibility using field survey gps data 120311.
and machine learning approaches. Fire 2 (3), 43. Pal, S., Paul, S., 2020. Assessing wetland habitat vulnerability in moribund Ganges delta
Ghosh, K.G., 2019. Spatial and temporal appraisal of drought jeopardy over the Gangetic using bivariate models and machine learning algorithms. Ecol. Indicat. 119, 106866.
West Bengal, eastern India. Geoenvironmental Disasters 6 (1), 1. Pandey, R.P., Pandey, A., Galkate, R.V., Byun, H., Mal, B.C., 2010. Integrating hydro-
Harmouzi, H., Nefeslioglu, H.A., Rouai, M., Sezer, E.A., Dekayir, A., Gokceoglu, C., 2019. meteorological and physiographic factors for ssessment of vulnerability to drought.
Landslide susceptibility mapping of the Mediterranean coastal one of Morocco Water Resour. Manag. 24, 4199–4217.
between Oued Laou and El Jebha using artificial neural networks (ANN). Arabian J. Pandey, S., Bhandari, H., 2006. Drought perpetuates poverty. Rice Today. April-June: 37.
Geosci. 12 (22), 696. Park, J., Sandberg, I.W., 1993. Approximation and radial basis function networks. Neural
Hong, H., Liu, J., Bui, D.T., Pradhan, B., Acharya, T.D., Pham, B.T., Zhu, A.X., Chen, W., Comput. 5, 305–316.
Ahmad, B.B., 2018a. Landslide susceptibility mapping using J48 decision tree with Pearson, K., 1897. Mathematical contributions to the theory of evolution. —on a form of
AdaBoost, bagging and rotation forest ensembles in the guangchang area (China). spurious correlation which may arise when indices are used in the measurement of
Catena 163, 399–413. organs. Proc. Roy. Soc. Lond. 60 (359–367), 489–498.
Hong, H., Liu, J., Zhu, A.X., 2020. Modeling landslide susceptibility using LogitBoost Pei, W., Fu, Q., Liu, D., Li, T.X., Cheng, K., Cui, S., 2018. Spatiotemporal analysis of the
alternating decision trees and forest by penalizing attributes with the bagging agricultural drought risk in Heilongjiang Province, China. Theor. Appl. Climatol. 133
ensemble. Sci. Total Environ. 718, 137231. (1–2), 151–164.
Hong, H., Tsangaratos, P., Ilia, I., Liu, J., Zhu, A.X., Chen, W., 2018b. Application of Pourghasemi, H.R., Gayen, A., Edalat, M., Zarafshar, M., Tiefenbacher, J.P., 2020. Is
fuzzy weight of evidence and data mining techniques in construction of flood multi-hazard mapping effective in assessing natural hazards and integrated
susceptibility map of Poyang County, China. Sci. Total Environ. 625, 575–588. watershed management? Geoscience Frontiers 11 (4), 1203–1217.

14
S. Saha et al. Journal of Cleaner Production 314 (2021) 128073

Rahmati, O., Tahmasebipour, N., Haghizadeh, A., Pourghasemi, H.R., Feizizadeh, B., Sudhira, H.S., Ramachandra, T.V., Subrahmanya, M.B., 2007. Bangalore. Cities 24 (5),
2017. Evaluating the influence of geo-environmental factors on gully erosion in a 379–390.
semi-arid region of Iran: an integrated framework. Sci. Total Environ. 579, 913–927. Talukdar, S., Ghose, B., Salam, R., Mahato, S., Pham, Q.B., Linh, N.T.T., Avand, M., 2020.
Rajendran, S., 2001. Drought in Karnataka: Need for long-term perspective. Econ. Polit. Flood susceptibility modeling in Teesta River basin, Bangladesh using novel
Wkly. 3423–3426. ensembles of bagging algorithms. Stoch. Environ. Res. Risk Assess. 34 (12),
Rasyid, A.R., Bhandary, N.P., Yatabe, R., 2016. Performance of frequency ratio and 2277–2300.
logistic regression model in creating GIS based landslides susceptibility map at Talukdar, S., Pal, S., 2020. Wetland habitat vulnerability of lower Punarbhaba river basin
Lompobattang Mountain, Indonesia. Geoenvironmental Disasters 3 (1), 19. of the uplifted Barind region of Indo-Bangladesh. Geocarto Int. 35 (8), 857–886.
Reddy, G., Prabhu, C.N., 2016. Natural Disaster Monitoring System – Karnataka Model, Tánago, I.G., Urquijo, J., Blauhut, V., Villarroya, F., De Stefano, L., 2016. Learning from
pp. 178–187. experience: a systematic review of assessments of vulnerability to drought. Nat.
Roy, J., Saha, S., 2020. Integration of artificial intelligence with meta classifiers for the Hazards 80 (2), 951–973.
gully erosion susceptibility assessment in Hinglo river basin, Eastern India. Adv. Termeh, S.V.R., Kornejady, A., Pourghasemi, H.R., Keesstra, S., 2018. Flood
Space Res. 67 (1), 316–333. susceptibility mapping using novel ensembles of adaptive neuro fuzzy nference
Rygel, L., O’sullivan, D., Yarnal, B., 2006. A method for constructing a social system and metaheuristic algorithms. Sci. Total Environ. 615, 438–451.
vulnerability index: an application to hurricane storm surges in a developed country. Thomas, T., Jaiswal, R.K., Galkate, R., Nayak, P.C., Ghosh, N.C., 2016. Drought
Mitig. Adapt. Strategies Glob. Change 11 (3), 741–764. indicators-based integrated assessment of drought vulnerability: a case study of
Saha, A., Saha, S., 2020. Comparing the Efficiency of Weight of Evidence, Support Vector Bundelkhand droughts in central India. Nat. Hazards 81 (3), 1627–1652.
Machine and Their Ensemble Approaches in Landslide Susceptibility Modelling: A Turner, B.L., Kasperson, R.E., Matson, P.A., McCarthy, J.J., Corell, R.W., Christensen, L.,
Study on Kurseong Region of Darjeeling Himalaya, India. Remote Sensing Eckley, N., Kasperson, J.X., Luers, A., Martello, M.L., Polsky, C., Pulsipher, A.,
Applications: Society and Environment, p. 100323. Schiller, A., 2003. A framework for vulnerability analysis in sustainability science.
Saha, S., Kundu, B., Paul, G.C., Mukherjee, K., Pradhan, B., Dikshit, A., Abdul Maulud, K. Proc. Natl. Acad. Sci. U.S.A. 100 (14), 8074–8079.
N., Alamri, A.M., 2021. Spatial assessment of drought vulnerability using fuzzy- Udmale, P., Ichikawa, Y., Manandhar, S., Ishidaira, H., Kiem, A.S., 2014. Farmers‫׳‬
analytical hierarchical process: a case study at the Indian state of Odisha. Geomatics, perception of drought impacts, local adaptation and administrative mitigation
Nat. Hazards Risk 12 (1), 123–153. measures in Maharashtra State, India. International Journal of Disaster Risk
Saha, S., Paul, G.C., Pradhan, B., Abdul Maulud, K.N., Alamri, A.M., 2020b. Integrating Reduction 10, 250–269.
multilayer perceptron neural nets with hybrid ensemble classifiers for deforestation Wilhite, D.A., Glantz, M.H., 1985. Understanding: the drought phenomenon: the role of
probability assessment in Eastern India. Geomatics, Nat. Hazards Risk 12 (1), 29–62. definitions. Water Int. 10 (3), 111–120.
Saha, S., Saha, M., Mukherjee, K., Arabameri, A., Ngo, P.T.T., Paul, G.C., 2020. Wu, Y., Ke, Y., Chen, Z., Liang, S., Zhao, H., Hong, H., 2020. Application of alternating
Predicting the deforestation probability using the binary logistic regression, random decision tree with AdaBoost and bagging ensembles for landslide susceptibility
forest, ensemble rotational forest and REPTree: a case study at the Gumani River mapping. Catena 187, 104396.
Basin, India. Sci. Total Environ. 139197. Yariyan, P., Janizadeh, S., Van Phong, T., Nguyen, H.D., Costache, R., Van Le, H.,
Selbesoglu, M.O., 2020. Prediction of tropospheric wet delay by an artificial neural Pham, B.T., Pradhan, B., Tiefenbacher, J.P., 2020. Improvement of best first decision
network model based on meteorological and GNSS data. Engineering Science and trees using bagging and dagging ensembles for flood probability mapping. Water
Technology, an International Journal 23 (5), 967–972. Resour. Manag. 34 (9), 3037–3053.
Shabbir, R., Ahmad, S.S., 2016. Water resource vulnerability assessment in Rawalpindi Yilmaz, I., Marschalko, M., Bednarik, M., Kaynar, O., Fojtova, L., 2011. Neural
and Islamabad, Pakistan using analytic hierarchy process (AHP). J. King Saud Univ. computing models for prediction of permeability coefficient of coarse grained soils.
Sci. 28 (4), 293–299. Neural Comput. Appl. https://doi.org/10.1007/s00521-011-0535-4.
Siebert, J., Sünnemann, M., Auge, H., Berger, S., Cesarz, S., Ciobanu, M., Eisenhauer, N., Yu, C., Chen, J., 2020. Landslide susceptibility mapping using the slope unit for
2019. The effects of drought and nutrient addition on soil organisms vary across southeastern helong city, jilin province, China: a comparison of ANN and svm.
taxonomic groups, but are constant across seasons. Sci. Rep. 9 (1), 1–12. Symmetry 12 (6), 1047.
Simelton, E., Fraser, E.D.G., Termansen, M., Forster, P.M., Dougill, A.J., 2009. Yuan, X.C., Wang, Q., Wang, K., Wang, B., Jin, J.L., Wei, Y.M., 2015. China’s regional
Typologies of crop-drought vulnerability: an empirical analysis of the socioeconomic vulnerability to drought and its mitigation strategies under climate change: data
factors that influence the sensitivity and resilience to drought of three major food envelopment analysis and analytic hierarchy process integrated approach. Mitig.
crops in China (1961e2001). Environ. Sci. Pol. 12 (4), 438e452. Adapt. Strategies Glob. Change 20 (3), 341–359.
Singha, P., Das, P., Talukdar, S., Pal, S., 2020. Modeling livelihood vulnerability in Zabihi, M., Pourghasemi, H.R., Motevalli, A., Zakeri, M.A., 2019. Gully erosion modeling
erosion and flooding induced river island in Ganges riparian corridor, India. Ecol. using GIS-based data mining techniques in Northern Iran: a comparison between
Indicat. 119, 106825. boosted regression tree and multivariate adaptive regression spline. In: Natural
Smakhtin, V.U., Hughes, D.A., 2007. Automated estimation and analyses of Hazards GIS-Based Spatial Modeling Using Data Mining Techniques. Springer, Cham,
meteorological drought characteristics from monthly rainfall data. Environ. Model. pp. 1–26.
Software 22 (6), 880–890. Zadeh, L.A., 1965. Fuzzy sets. Inf. Contr. 8 (3), 338–353.
Sönmez, F.K., Koemuescue, A.U., Erkan, A., Turgu, E., 2005. An analysis of spatial and Zhang, Q., Sun, P., Li, J., Xiao, M., Singh, V.P., 2015. Assessment of drought vulnerability
temporal dimension of drought vulnerability in Turkey using the standardized of the tarim river basin, xinjiang, China. Theor. Appl. Climatol. 121 (1–2), 337–347.
precipitation index. Nat. Hazards 35 (2), 243–264. Zhao, G., Pang, B., Xu, Z., Peng, D., Xu, L., 2019. Assessment of urban flood susceptibility
Sridevi, D., Jyotishi, A., Mahapatra, S., Jagadeesh, G., Bedamatta, S., 2014. Climate using semi-supervised machine learning model. Sci. Total Environ. 659, 940–949.
Change Vulnerability in Agriculture Sector: Indexing and Mapping of Four Southern
Indian States.

15