Science of the Total Environment 801 (2021) 149797

Contents lists available at ScienceDirect

Science of the Total Environment

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

Interpretable and explainable AI (XAI) model for spatial

drought prediction
Abhirup Dikshit a, Biswajeet Pradhan a,b,c,⁎
a
Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS), Faculty of Engineering and IT, University of Technology Sydney, NSW 2007, Australia
b
Department of Energy and Mineral Resources Engineering, Sejong University, Choongmu-gwan, 209 Neungdong-ro, Gwangjin-gu, Seoul 05006, Republic of Korea
c
Earth Observation Center, Institute of Climate Change, Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor, Malaysia

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• An explainable machine learning model

was used for first time for spatial
drought prediction.
• The model showed an accuracy of over
85% for all the five study sites.
• SHAP plots were developed to examine
predictor interactions for various
drought conditions.
• Plots illustrated the variation in predic-
tor relationships before and after Mil-
lennium Drought.

a r t i c l e i n f o a b s t r a c t

Article history: Accurate prediction of any type of natural hazard is a challenging task. Of all the various hazards, drought predic-
Received 5 July 2021 tion is challenging as it lacks a universal definition and is getting adverse with climate change impacting drought
Received in revised form 16 August 2021 events both spatially and temporally. The problem becomes more complex as drought occurrence is dependent
Accepted 16 August 2021
on a multitude of factors ranging from hydro-meteorological to climatic variables. A paradigm shift happened in
Available online 21 August 2021
this field when it was found that the inclusion of climatic variables in the data-driven prediction model improves
Editor: Fernando A.L. Pacheco the accuracy. However, this understanding has been primarily using statistical metrics used to measure the
model accuracy. The present work tries to explore this finding using an explainable artificial intelligence (XAI)
model. The explainable deep learning model development and comparative analysis were performed using
Keywords: known understandings drawn from physical-based models. The work also tries to explore how the model
Drought prediction achieves specific results at different spatio-temporal intervals, enabling us to understand the local interactions
Deep learning among the predictors for different drought conditions and drought periods. The drought index used in the
Hybrid model study is Standard Precipitation Index (SPI) at 12 month scales applied for five different regions in New South
Explainable AI
Wales, Australia, with the explainable algorithm being SHapley Additive exPlanations (SHAP). The conclusions
GIS
drawn from SHAP plots depict the importance of climatic variables at a monthly scale and varying ranges of an-
Australia
nual scale. We observe that the results obtained from SHAP align with the physical model interpretations, thus
suggesting the need to add climatic variables as predictors in the prediction model.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction
⁎ Corresponding author at: Centre for Advanced Modelling and Geospatial Information
Systems (CAMGIS), Faculty of Engineering and IT, University of Technology Sydney, NSW
2007, Australia. Droughts are one of the most severe, complex and devastating natu-
E-mail address: [email protected] (B. Pradhan). ral hazards that can last for several years or even decades. It is also one

https://doi.org/10.1016/j.scitotenv.2021.149797
0048-9697/© 2021 Elsevier B.V. All rights reserved.
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

of the least understood and the most costly natural disasters (Wang Therefore, in this paper, we investigate the use of a hybrid model involv-
et al., 2015). As per the Emergency Events Database (EM-DAT), 2020 ing LSTM and CNN to predict meteorological drought at different geo-
compilation, over 300 global drought occurrences in the 21st Century, graphic coordinates.
affecting over a billion people. Australia is one of the most drought af- The hindrance towards using any machine learning model in real-
fected regions and is also the driest inhabited continent which has world scenarios by drought management planners is the lack of trans-
seen multiple and severe drought events (Dikshit et al., 2020c). parency, explainability, and trust, as these are considered being black
The basic definition of drought is water deficiency compared to nat- box models (Chakraborty et al., 2021; Dikshit and Pradhan, 2021). The
ural circumstances (West et al., 2019). However, in reality, the defini- decision-makers tend to use physical-models even though several stud-
tion is much more complicated. Lloyd-Hughes (2014) suggested that a ies have found ML models outperforming them in terms of accuracy.
universally accepted drought definition is unavailable and may never The Bureau of Meteorology (BoM) uses the Predictive Ocean-
be possible. Therefore, based on drought impacts, it has been catego- Atmosphere Model for Australia (POAMA), which provides seasonal to
rized into different types: meteorological, hydrological, agricultural, inter-annual seasonal rainfall forecasts based on a physical-based
ground-water, ecological, environmental and socio-economic model. Abbot and Marohasy (2014) used ANN to forecast rainfall in
(Crausbay et al., 2017; Mishra and Singh, 2010; Vicente-Serrano et al., Queensland and found their results to be better performing than the
2020). To determine drought, drought indices have been developed physical-based model based forecasts. Therefore, it is of utmost impor-
which quantify single or multiple drought affecting variables in a stan- tance that the model outputs be explained and interpreted. The emer-
dardized or threshold-based way to represent different drought charac- gence of new explainable algorithms, like SHapley Additive
teristics. In the case of meteorological drought, which is the present exPlanations (SHAP), can not only help to understand model outputs
study's focus, the Standardized Precipitation Index (SPI) developed by but also change perception of using ML based models for decision-
McKee et al. (1993) is most commonly used. The index is robust, statis- making purposes.
tically simple to compute, independent of climatic factors, and has also The novelty of the work lies in the application of an explainable ma-
been accepted globally as a drought monitoring index by the World chine learning framework to predict drought and to identify which var-
Meteorological Organization (WMO) (Hayes et al., 2011). iables are contributing to the predictive model and their relative
The occurrence of a drought event is based on multiple factors, importance. This is for the first time where the attempt has been
which are region specific, and therefore, a universal model cannot be made to explain how a machine learning model achieves a specific re-
developed (Deo et al., 2017). The focus of the study is New South sult at different temporal ranges for various locations under various
Wales, in Australia, where drought problems have been aggravated by drought scenarios. Hence, the present study also looks at understanding
climatic influences led by strong El-Nino influences (Deo et al., 2017; the model outputs through the use of different SHAP plots. In summary,
Dikshit et al., 2021a; Verdon-Kidd and Kiem, 2009). There has been a the objectives of the study are: i) Use of a CNN-BiLSTM based model to
deficiency of rainfall and an increase in temperature over the last predict meteorological drought index (SPI); ii) Analyzing the individual
50 years (Dikshit et al., 2020b). The longest drought (Millennium predictions using Shapley outputs to understand the model-predictor
Drought) (1997–2009) was found to be a resultant of a combination relationship for different drought conditions and iii) Examining the spa-
of Southern Annular Mode (SAM) and anomalies in the Indian Ocean tial variation of model outputs in different geographic conditions.
(Ummenhofer et al., 2009). A strong La Niña event led to an unprece-
dented rainfall event, leading to the end of the millennium drought 2. Study area
(Boening et al., 2012; Ummenhofer et al., 2009). Previous studies have
also found that the relationship between climatic drivers and rainfall The area of focus is the state of New South Wales (NSW) in Eastern
in Australia is the highest in the world (Kirono et al., 2010). Australia, which covers an area of over 800,000 km2, with more than
The forecasting model developed for drought has seen the rise of three-quarters of land under agriculture (Smart, 2016). The region is
data-driven models (Hao et al., 2018; Mishra and Singh, 2011). Different also home to the Murray-Darling Basin (MDB) which is the most impor-
data-driven models, especially neural networks, have achieved promising tant agricultural area and produces nearly 33% of the national food sup-
results and detailed information about the various modelling studies with ply (Smart, 2016). The region has a history of major drought events
their advantages and disadvantages have been compiled by various re- (Leblanc et al., 2012) and the recent bushfire event in 2019 has also
view articles (Fung et al., 2020; Hao et al., 2018). In the Australian con- been attributed to drought like conditions combined with dry vegeta-
text, drought forecasting using different models has been applied, tion and increase in temperature (Nolan et al., 2020). The minor
providing satisfactory results. Like, Deo et al. (2017) used three machine drought events have also immensely contributed to huge economic
learning (ML) models to forecast monthly SPI in five different locations losses, with estimates suggesting a loss of over 8 Billion$ in
and found periodicity to be an important factor. Dikshit et al. (2020b, 2017–2020 excluding the bushfire event (Dikshit et al., 2021b). The
2020c) predicted Standard Precipitation Evaporation Index (SPEI) at droughts in the region are a resultant of several factors, including but
different time scales using artificial neural network (ANN) and random not limited to climatic influences led by anthropogenic climate change,
forests model and suggested that the examination of forecasted results resulting in erratic rainfall and temperature patterns.
should be based on drought characteristics rather than statistical metrics. The climate in the region responds to oceanic phases, defined by
Barua et al. (2012) used ANN to forecast Non-linear Aggregated Index multiple climatic drivers such as the Southern Oscillation Index (SOI),
(NADI) index in Victoria and compared the results with stochastic Pacific Decadal Oscillation Index (PDO), Indian Ocean Dipole (IOD)
models, and found ANN to be performing better. and Southern Annular Mode (SAM). The sea surface temperature
Neural networks are incapable of dealing with non-stationarities in (SST) based indices are Nino 3.0, Nino 3.4 and Nino 4.0 (Nicholls,
drought estimations and suffer from overfitting due to lag components 2004; Ummenhofer et al., 2009; Van Dijk et al., 2013). Past studies
involved in time series data (Dikshit et al., 2021a). The emergence of have also shown that changes in IOD and PDO phases are correlated
deep learning models has been found to effectively solve this problem with drought occurrences (Deo et al., 2017), with the rainfall patterns
(Dikshit et al., 2021a, 2021b; Poornima and Pushpalatha, 2019). These in MDB is influenced by changes in SOI, IOD and PDO phases
studies have looked upon the use of a dynamic recurrent neural net- (McGowan et al., 2009).
work, Long Short-term Memory model (LSTM) and have achieved satis- The present work studies five different sites in the NSW state to test
factory results. The studies on deep learning for drought forecasting are the capabilities of the model under different geographical conditions.
in infancy, with very limited use cases. Other sophisticated deep learn- The monthly rainfall data was collected from Scientific Information for
ing models like Convolutional Neural Network (CNN) based have not Land Owners (SILO). Fig. 1 depicts the location, with Table 1 providing
been explored and can be effectively used for time-series forecasting. some key statistics about rainfall at different sites.

2
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 1. Location of the five study sites.

2.1. Standardized precipitation index 1967) multiple drought characteristics can be determined. For example,
drought characteristics, such as drought onset, which indicates drought
The Standardized Precipitation Index (SPI) was conceptualized by initiation (SPI < 0), drought end which shows drought end (SPI > 0),
(McKee et al., 1993), is a non-linear probabilistic-based index derived drought duration which is the time period between drought onset and
by standardizing monthly rainfall distribution, using a Gamma or end. Fig. 2 shows these drought conditions for Barraba station based
Pearson type III distribution. The index can be calculated at different on SPI-12 values from July 2001 – June 2004. As the figure suggests, dur-
time-scales, which indicates different drought types. The detailed ing the specified time period, January 2003 had the peak intensity with
understanding of the calculation of the SPI index has been explained a SPI value of −2.56. Similarly, drought onset started in January 2002,
by Belayneh et al. (2014) and Deo et al. (2017). In this study, a and ending in December 2003.
Gamma probability distribution was used to calculate SPI using ‘SPEI’
package in R (Vicente-Serrano et al., 2010). The values range from −2 3. Models and algorithms used
to +2, with negative values indicating drought and can be categorized
under drought categories as depicted in Table 2 (Morid et al., 2007). 3.1. Long-short term memory model
Based on the values and using the running sum technique (Yevjevich,
LSTM is a dynamic recurrent neural network (RNN), which has a
better control over gradient flow, overcoming the RNN's problem of
vanishing gradient (Hochreiter and Schmidhuber, 1997). It ensures bet-
Table 1 ter preservation of long-term dependencies by filtering out redundant
Study sites and their basic statistical information.
or misleading information. LSTM does this with the help of memory
Location Geographical Elevation (m) Rainfall statistics
coordinates (1901–2018)(mm/month)
Table 2
Latitude Longitude Mean Standard deviation Drought categories based on negative SPI values.
Bathurst 33.43°S 149.56°E 713 52.35 38.04
SPI values Drought categories
Barraba 30.38°S 150.61°E 499 56.98 47.84
Peak Hill 32.72°S 148.19°E 285 46.67 40.9 −1.49-−1 Moderately dry
Collarenebri 29.55°S 148.55°E 145 41.63 45.22 −1.99-−1.5 Severely dry
Yamba 29.43°S 153.36°E 27.4 121.62 101.18 ≤−2 Extremely dry

3
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 4. Schematic illustration of CNN-BiLSTM architecture used in this study.

of both the models. A schematic diagram of the architecture is presented

in Fig. 4.
The performance of the model is dependent on the choice of
hyperparameters (Choudhary et al., 2020). This is more important for
RNN based models, which has a more complex architecture than con-
ventional neural networks and thus the choice of hyperparameters af-
fects the performance of the model. A general approach towards the
Fig. 2. SPI-12 (orange lines) with monthly rainfall values (blue bars) values and
corresponding drought conditions for Barraba station (July 2001 – June 2004). (For
choice of hyperparameters for both the algorithms involves the follow-
interpretation of the references to colour in this figure legend, the reader is referred to ing steps: i) Choice of initial hyper parameter set; ii) learning rate; iii)
the web version of this article.) batch size and number of epochs; iv) selecting best optimizer and acti-
vation function and lastly, v) optimum number of layers and neurons.

3.3. Shapley additive explanations (SHAP)

cell, input (i), output (ot) and forget gates (ft). Each of the LSTM units
works in tandem, which is unlike vanilla RNN that has a single hidden SHAP was first introduced in game theory to determine the contri-
layer. The details about the use of LSTM architecture for drought fore- bution of an individual player in a collaborative game (Shapley, 1953).
casting have been explained in detail in (Dikshit et al., 2021a, 2021b; The idea was to distribute the total gain among players based on their
Poornima and Pushpalatha, 2019). contributions towards the outcome. The SHAP values provided a solu-
A Bidirectional LSTM (Bi-LSTM) comprises of two LSTM compo- tion to the problem of a fair reward to every player and assigning a
nents, one is the forward LSTM and the other is backward LSTM. unique value determined by local accuracy, consistency and null effect
These two components concatenate to form the output. The motive (Shapley, 1953). The recent development of machine learning algo-
behind applying a backward LSTM is to provide a path for the net- rithms by Lundberg and Lee (2017) has opened up new avenues to-
work to examine the future values. This helps the system in learning wards understanding the model outputs, thus providing more
the finer dependencies in a dataset which has periodic aspects which transparency to the usually considered black box models.
can be evidently seen in a majority of the time series forecasting The shapley value is calculated based on the average of the marginal
problems (Fig. 3). contribution across all the possible permutations of the features. The
mathematical expression of the classical SHAP value is as below:
3.2. CNN-BiLSTM model
jSj!ðn−jSj−1Þ!
∅i ¼ ∑S⊆Nfig ½vðS∪figÞ−vðSÞ ð1Þ
n!
Convolutional neural network (CNN) has been proven effective in
time-series problems (Krizhevsky et al., 2012). It mainly consists of where, ∅_i is the contribution of feature i, N is the set of all the features,
three main layers: a) Convolutional layers, which reads input data se- n is the number of features in N, S is the subset of N which contains the
quences and automatically extract relevant features, b) Pooling layers feature i and v(N) is the base value meaning the predicted outcome for
which extracts important features and focuses on the important vari- each feature in N without knowing the feature values.
ables and c) Fully connected layers that interprets the internal represen- The model outcome for each observation is estimated by summing
tation and output a vector representing multiple time steps (Brownlee, up the SHAP value of each feature for that observation. Therefore, the
2016). explanation model is formulated as below:
The key benefits of the proposed hybrid model are the automatic
M
feature learning and the ability of the model to output a multi-step vec- g ðz0 Þ ¼ ∅0 þ ∑i¼1 ∅i z0i , ð2Þ
tor directly. The CNN architecture can extract and learn localised fea-
tures, whereas Bi-LSTM can capture the variations in long-term where, z′ ϵ {0, 1}M and M is the number of features and ∅_i can be ob-
dependencies (Dikshit et al., 2020a). Therefore, the present work uses tained from the eq. 2. SHAP provides multiple AI model explainers.
a hybrid model of CNN and Bi-LSTM, thereby combining salient features The details about different model explainers are beyond the scope of

Fig. 3. (a) LSTM unit input (i), output (ot) and forget gates (ft); (b) Bi-LSTM architecture with one forward and one backward LSTM layer.

4
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

this study. Molnar (2020) provides a detailed explanation about differ- Table 3
ent explainers and the corresponding plots with different use cases. In Statistical metrics of the predicted results for the five studied areas.

the present study, we use a Deep Explainer designed for deep learning Location R2 RMSE MAE
algorithms. Bathurst 0.90 0.30 0.25
Traditionally, research works involving ML based models for Barraba 0.90 0.30 0.23
drought forecasting, use a variable importance plot to show the most Peak Hill 0.89 0.31 0.24
and least important predictors (Dikshit et al., 2020c; Khan et al., Collarenebri 0.91 0.33 0.27
Yamba 0.88 0.33 0.25
2020). Such variable importance plots fails to explain key aspects like
the interaction among the variables to reach a specific model output,
which could be for the entire testing period or for a specific time period.
Therefore, we explore the individual and collective force plots to under-
stand how the variables affect the model outcomes in different drought Previous studies have also reported similar or sometimes even bet-
conditions at various locations. Thereafter, we compare the SHAP find- ter results based on different drought indices and predictors. To under-
ings with the previous studies suggesting the causative factor of stand the effect of variables, ablation study has often been used (Dikshit
droughts and the influence of climate change on drought occurrences et al., 2020b; Khan et al., 2020), wherein predictors are removed or
(Cai et al., 2014). added gradually and the corresponding changes in statistical metrics
is noted, thereby finding the best possible combination. However,
3.4. Model development such processes can be time-consuming and is highly dependent on the
test dataset with no adequate understanding at individual time steps.
The review by Anshuka et al. (2019) on drought forecasting stud- However, SHAP provides local interpretability by enabling us to identify
ies using SPI as the drought index suggested that SPI at longer time the impact of each variable on SPI prediction for individual location. This
scales are better forecasted, as the values are smoother. However, capability allows us to analyse the results in more detail, which leads us
the selection of time scale is dependent on the study area and prior to a deeper interpretation of the prediction model result for different
tests must be done for selecting the relevant time scale. The present drought conditions.
work uses SPI-12 as the drought index and large-scale climatic influ-
ences as predictors in the prediction model. The predictive capacity 4.1. SHAP force plots
of the model is checked using three statistical metrics, Coefficient
of Determination (R 2 ), Root-Mean-Square-Error Method (RMSE) The individual SHAP force plot helps us to understand how the
and Mean-Absolute-Error (MAE). The mathematical formulae of variables interact with each other to reach the predicted target vari-
these metrics are: able (Matin and Pradhan, 2021). These plots have three important
characteristics i.e.: i) output value: predicted value of the individual

̿

R2 ¼ ∑k−1
m
xi −xi =∑m ðx −x Þ2 ð3Þ observation; ii) base value: average predicted observations of the
k¼1 i i
test dataset; and iii) colours: variables in red colour push the predic-
tion higher whereas the blue colour pushes the prediction lower.
̿ m
xi ¼ 1=m∑k¼1 xi , ð4Þ However, as SPI-12 values have both positive and negative values,
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi it doesn't follow the conventional way to analyse SHAP plots. In
m cases where the predicted value is less than the base value, the
∑k¼1 ðxi −xi Þ2
RMSE ¼ , ð5Þ blue colour pushes the prediction higher, and the red colour pushes
m
the prediction lower, and vice-versa.
m ∣xi −xi ∣ Therefore, in order to interpret the model outcome for different
MAE ¼ ∑k¼1 , ð6Þ drought condition SHAP has been employed for different drought sce-
m
narios: i) extreme dry; ii) severely dry; iii) moderately dry; and iv)
̿
where, xi is the mean value, xi and xi are observed and forecasted values non-drought conditions at different sites to understand the interactions
with m being the number of samples. among variables towards model output. Table 4 provides details about
The layers employ Rectified Linear Unit (ReLU) as their non- the drought conditions, along with their actual and predicted values.
linear activation function and the final dense layer has a single pre- The model explanation was generated for all the sites using SHAP
dictor neuron. Adaptive momentum (Adam) is the optimiser and, force plots.
given the need for regression, the loss function was ‘mean absolute
error’. The deep learning algorithm was applied with Keras 2.0 4.1.1. Bathurst
(Francois, 2016) using TensorFlow 2.0 as the back end. The data For Bathurst station, there was only one time period when SPI was
was split into training (1901–1991), validation (1992–2001) and less than −2, which was January 2007, and as Fig. 6(a) shows that rain-
testing (2002–2018). To capture the high-level features, we use a fil- fall was a major contributor with SAM and Nino 3.4 being the other con-
ter size of 3 × 3 with stride length of 2. The high level features are ag- tributors increasing the model output, whereas IOD slightly decreasing
gregated using a max pooling layer. The batch size was 32 and it. For severely dry condition (Fig. 6b), climate indices excluding SOI de-
epochs were 500, with the sequence length being 12. The experi- creased the model output with rainfall increasing the model output. In
ments were conducted on four Nvidia V100s with 16GB of GPU the case of moderate dry condition (Fig. 6c), IOD and SOI decreased
VRAM. the model output, with other predictors like rainfall, Nino 4, Nino 3.4
and SAM increased the model output. For non-drought conditions, rain-
4. Results fall, Nino 3, Nino 3.4, SOI and IOD increased and SAM decreased the
model output.
This section discusses the model results based on statistical metrics The first two figures (Fig. 6a; b) exhibit drought conditions during
and examines the model outcomes using Shapley values. The statistical the Millennium Drought period (1997–2009), which was the worst
metrics for all the locations are presented in Table 3. drought occurrence in Australia in the recent history. Different works
The variation between actual and predicted SPI-12 values is illus- have analyzed this drought period. Therefore, we look at SHAP force
trated in Fig. 5. As the figures show, the model was able to predict the plots of rainfall and IOD for the entire Millennium drought period and
SPI-12 values with R2 higher than 0.85 at all the locations. examine the variation.

5
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 5. Variation between actual and predicted values for all the sites (a) Bathurst; (b) Barraba; (c) Peak Hill; (d) Collarenberi; and (e) Yamba during the testing period.

The SHAP values represent the additive effects of each feature, correlated would not reveal the effect of individual predictors on the
i.e., the cumulative sum of effects of individual predictors will be the prediction capability. In Fig. 7, we observe that the additive contribution
final shapley value. Due to this additive feature, predictors which are of rainfall is significantly higher than IOD across the range of values in

6
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Table 4
Drought conditions and time periods analyzed using SHAP individual force plots.

Station name Drought conditions Figures Month Actual values Predicted values

Bathurst Extreme dry Fig. 6a January 2007 −2.11 −2.01

Severely dry Fig. 6b March 2007 −1.73 −1.57
Moderately dry Fig. 6c October 2017 −1.39 −1.07
Non drought Fig. 6d December 2010 1.50 1.69
Barraba Extreme dry Fig. 8a February 2003 −2.20 −2.11
Severely dry Fig. 8b February 2007 −1.67 −1.23
Moderately dry Fig. 8c September 2018 −1.19 −1.14
Non drought Fig. 8d February 2012 1.74 1.14
Peak Hill Extreme dry Fig. 9a March 2007 −2.20 −1.93
Severely dry Fig. 9b June 2007 −1.58 −1.45
Moderately dry Fig. 9c March 2013 −1.12 −1.17
Non drought Fig. 9d February 2017 1.72 1.75
Collarenebri Extreme dry – – –
Severely dry Fig. 10a September 2013 −1.59 −1.84
Moderately dry Fig. 10b March 2014 −1.27 −1.10
Non drought Fig. 10c December 2010 1.57 1.35
Yamba Extreme dry Fig. 11a April 2005 −2.04 −1.70
Severely dry Fig. 11b October 2004 −1.65 −1.34
Moderately dry Fig. 11c August 2004 −1.46 −1.44
Non drought Fig. 11d August 2011 2.50 2.40

the abscissa. The values on abscissa represent the individual additive occurrence (Bureau of Meteorology, Australia). This was also the pe-
contribution of each feature to the cumulative shapley value. However, riod which recorded the highest temperature compared to the mean
as rainfall and IOD are correlated (Ashok et al., 2003), the underlying as- of mid-20th century (Dey et al., 2019). Fig. 8c depicts a moderate dry
sumption of the variables being independent is no longer valid. This condition in September 2018, and as expected rainfall was the major
leads to dependence explosion, which inflates the overall contribution contributing factor. Among the climatic variables, SOI and nino4 in-
of rainfall and vice-versa. Therefore, we see a major contribution of creases the model output. Other variables like SAM, PDO and IOD, de-
rainfall in events that were primarily dependent on IOD. crease the model output.

4.1.2. Barraba 4.1.3. Peak Hill

Similar to Bathurst, Fig. 8(a, b) exhibits drought conditions during Fig. 9 (a, b) depicts different drought conditions during the latter half
the Millennium drought and as expected rainfall was the most con- of the Millennium drought and the predictor relationship is similar to
tributory factor and climatic variables affected either positively or other stations with minor changes. However, these changes are minis-
negatively depending on the time-scale and drought condition. The cule and the variation among the predictors towards prediction cannot
station belongs to the MDB, which had seen substantially low rainfall be quantified. By mid-2013, more than 15% of NSW was under drought
during 2017–2019, with no clear historical precedent for such an and within the next 9 months, 62% of the state had received no

Fig. 6. Individual force plots for Bathurst station for different drought and non-drought conditions.

7
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 7. Collective force plots of individual predictors (a) rainfall; (b) IOD during the millennium drought.

significant rainfall (Bureau of Meteorology, Australia, 2020). Fig. 9c except Nino 4 along with climatic indices decreasing the model output
shows a moderately dry condition for March 2013, just prior to the and Nino 4 increasing the model output.
more severe drought conditions prevailing in the state. In both the
time periods, climatic variables played an important role in drought pre- 4.1.4. Collarenebri
diction. In case of non-drought condition (Fig. 9d), for February 2017 For the Collarenebri station, there was no instance of extreme dry
also shows the importance of climatic variables with Nino indices condition during the testing period. For other drought conditions

Fig. 8. Individual force plots for Barraba station for different drought and non-drought conditions.

8
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 9. Individual force plots for Peak Hill station for different drought and non-drought conditions.

(Fig. 10a, b) the examined time period is during the 2013–2014 drought cases caused flooding (Cai and Van Rensch, 2012; Verdon-Kidd and
period, which was a relatively short duration, also known as flash Kiem, 2009). Fig. 11d explains the interactions in August 2011, depicting
droughts. In both the cases, SOI decreases the model output, whereas the Nino indices improving the model outcome, with SAM and PDO
other variables increase the model output. In case of non-drought con- decreasing the model outcome.
dition (Fig. 10c), for December 2010, which also marked the end of
Millennium Drought, shows rainfall and other climatic variables to be 5. Discussions
increasing the model output, whereas variables like SOI and SAM de-
creasing the model output. This paper uses a hybrid deep learning model, a combination of CNN
and LSTM to predict drought index (SPI-12) for different stations in
4.1.5. Yamba NSW. The monthly rainfall data was collected from SILO ranging from
Fig. 11(a–c) depict the model-predictor relationship for different 1901 to 2018, of which 1901–1991 was used for training, 1992–2001
drought conditions during the Millennium drought period and as for validation and the remaining (2002–2018) as testing dataset. The
explained in the previous section, it follows the same pattern, with results depict that the model achieved R2 value of more than 0.85, with-
some minor changes suggesting little spatial variation. The end of the out performing any pre-processing, normalization or involving complex
Millennium drought in early 2010 was due to La Niña event combined architectures. This was done as there are developmental challenges
with anomalously warm SSTs, leading to heavy rainfall and in some with SHAP, and as the model complexity increases, the Deep Explainer

Fig. 10. Individual force plots for Collarenebri station for different drought and non-drought conditions.

9
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Fig. 11. Individual force plots for Yamba station for different drought and non-drought conditions.

doesn't perform adequately. Therefore, due to the current prevailing non-drought conditions, like the end of Millennium drought which
challenges of SHAP for deep learning model, simple architectures can was caused to a La Niña event (Cai and Van Rensch, 2012) was also ad-
be used to understand if the future direction in the field should be equately captured by SHAP plots showing the importance of nino indi-
model agnostic or data agnostic. It is important to understand that the ces.
framework of SHAP values only provides a co-operative game theoretic When examining different drought periods, the results suggest that
approach towards understanding the contribution of different input the individual SHAP force plots can provide a broad understanding of
features towards the final model outcome. Rudin (2019) mentioned the possible reasons for model outputs. For example, Nguyen et al.
that the notion about explainable ML models being not as accurate as (2021) in their study on 2019 flash droughts showed the importance
complex deep learning models is a myth. The work also mentioned of large-scale climatic drivers on flash droughts. Although, it would be
that a trade-off between model predictive accuracy and explainability premature to say if SHAP can fully explain the concept of flash drought,
seems to be refraining end-users from developing explainable based as at present there is no published works on the use of ML models to
models. The work presented here eliminates such beliefs by applying forecast flash drought, with a majority of the works looking to define
an explainable ML based model for drought based applications. It has and adequately capture it (Otkin et al., 2018; Pendergrass et al., 2020).
the capability to accurately portray information to experts and end- However, as and when more works are published, researchers could
users about how the model makes predictions and should be adopted also look towards the use of SHAP to understand the effect of climatic
for broader hydro-geologic applications. Therefore, this work aims to variables on flash droughts.
understand SHAP plots for different drought conditions in different re- As these plots show, it is important to understand the drought con-
gions. The real-world significance of these has been compared with ditions as well as the time-period when drought events occurred. When
the individual and collective SHAP force plots. examining the spatial variation, there exists some difference when ana-
In accordance, the paper introduces the use of SHAP force plots to lyzing from drought condition point of view. However, the spatial vari-
generate the explanation for the CNN-BiLSTM model using both individ- ation among the contribution of specific predictors in different drought
ual and cumulative time-series data and understands how the predic- conditions is not very significant. This could be due to a shorter lead
tors affect the model output. We look at three different drought and time as well as smaller test dataset. One can look to generalize the fea-
non-drought conditions for a specific time period using the test dataset. ture interaction based on either time periods (seasonal, decadal or
The plots explain how the predictor relationship helped achieve the drought event) or drought conditions, and more works need to be con-
prediction value, i.e., which predictors increased or decreased model ducted to understand the variation in different case scenarios.
output. Such an analysis could explain the intricacies for individual It is important to note that the SHAP algorithm was specifically not
plots, which would be important for real-world scenarios when future designed for time-series problems, with the development of Deep
predictor data is unavailable. To confirm the findings, we compared Explainer still being in its nascent stage (Molnar, 2020). The Deep
the SHAP plots for the Millennium drought period and found IOD to Explainer model of SHAP is able to take any black-box model as input
be an important factor which was also found responsible for the to provide the significance of its input features to its output. However,
drought event. The findings show the importance of climatic variables this generalization also implies that SHAP inherently does not under-
in different time-periods and different drought conditions, thus necessi- stand the inner workings of a model. Consequently, when employed
tating its use in drought prediction model. Ummenhofer et al. (2009), in handling complex data such as time series that utilizes temporal fea-
suggested that the ENSO variation which has been usually considered tures, Deep Explainer is not able to understand the temporal implica-
for droughts in eastern Australia cannot fully explain such drought con- tions of the features and needs defined aggregation from the user's
dition. Their study used monthly rainfall from SILO dataset and climatic side. Here, we attempt to use SHAP to understand the interaction be-
indices and found IOD to be a predominant factor. However, both the tween predictors based on the additive property of SHAP. The force
figures at individual time step showed different results. Similarly, for plots depicted here provide a general understanding of the variable

10
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

interaction and do not reflect the exact value. However, the SHAP values Belayneh, A., Adamowski, J., Khalil, B., Ozga-Zielinski, B., 2014. Long-term SPI drought
forecasting in the Awash River basin in Ethiopia using wavelet neural network and
are correlated to the actual prediction and thus are able to provide valu- wavelet support vector regression models. J. Hydrol. 508, 418–429.
able information. Boening, C., Willis, J.K., Landerer, F.W., Nerem, R.S., Fasullo, J., 2012. The 2011 La Niña: so
strong, the oceans fell. Geophys. Res. Lett. 39.
Brownlee, J., 2016. Supervised and unsupervised machine learning algorithms. Mach.
6. Conclusions
Learn. Mastery 16.
Cai, W., Purich, A., Cowan, T., van Rensch, P., Weller, E., 2014. Did climate change–induced
Droughts are one of the most challenging natural hazards to be pre- rainfall trends contribute to the Australian Millennium Drought? J. Clim. 27,
dicted owing to multiple predictors and non-linearity between them. 3145–3168.
Cai, W., Van Rensch, P., 2012. The 2011 Southeast Queensland extreme summer rainfall: a
The main objective of the paper was to provide more transparency confirmation of a negative Pacific decadal oscillation phase? Geophys. Res. Lett. 39.
and explainability when using AI based models in the field of drought Chakraborty, D., Basagaoglu, H., Winterle, J., 2021. Interpretable vs. noninterpretable ma-
forecasting. This would help in building trust among the government chine learning models for data-driven hydro-climatological process modelling. Ex-
decision-makers and planners, which have for long used physical pert Syst. Appl. 170, 114498.
Choudhary, N., Rao, N., Katariya, S., Subbian, K., Reddy, C.K., 2020. Self-supervised Hyper-
based models for examining future drought scenario. The main findings boloid Representations From Logical Queries Over Knowledge Graphs arXiv preprint
from the study are as: arXiv:2012.13023.
Crausbay, S.D., Ramirez, A.R., Carter, S.L., Cross, M.S., Hall, K.R., Bathke, D.J., Betancourt, J.L.,
Colt, S., Cravens, A.E., Dalton, M.S., 2017. Defining ecological drought for the twenty-
• Deep learning can prove to be an effective approach to predict mete-
first century. Bull. Am. Meteorol. Soc. 98, 2543–2550.
orological drought and its wide acceptability will need to be further Deo, R.C., Kisi, O., Singh, V.P., 2017. Drought forecasting in eastern Australia using multi-
tested based on different drought types and drought indices. variate adaptive regression spline, least square support vector machine and M5Tree
• The use of explainable algorithms like SHAP is important to build model. Atmos. Res. 184, 149–175.
Dey, R., Lewis, S.C., Arblaster, J.M., Abram, N.J., 2019. A review of past and projected
more trust and provide transparency towards the use of ML models, changes in Australia's rainfall. Wiley Interdiscip. Rev. Clim. Chang. 10, e577.
thereby embracing data-driven models for its use in drought early Dikshit, A., Pradhan, B., Alamri, A.M., 2020a. Pathways and challenges of the application of
warning system. artificial intelligence to geohazards modelling. Gondwana Res. https://doi.org/10.
1016/j.gr.2020.08.007.
• Using SHAP explainer, the study found the importance of climatic var-
Dikshit, A., Pradhan, B., Alamri, A.M., 2020b. Short-term spatio-temporal drought forecast-
iables as predictors in drought prediction model across different ing using random forests model at New South Wales, Australia. Appl. Sci. 10, 4254.
drought conditions and drought events, and confirming its inclusion Dikshit, A., Pradhan, B., Alamri, A.M., 2020c. Temporal hydrological drought index fore-
in the prediction model. casting for New South Wales, Australia using machine learning approaches. Atmo-
sphere 11, 585.
• Future works should look towards analyzing the SHAP plots for long
Dikshit, A., Pradhan, B., Alamri, A.M., 2021a. Long lead time drought forecasting using
lead time forecasting along with examining other additive SHAP lagged climate variables and a stacked long short-term memory model. Sci. Total En-
properties. viron. 755, 142638.
Dikshit, A., Pradhan, B., Huete, A., 2021b. An improved SPEI drought forecasting approach
using the long short-term memory neural network. J. Environ. Manag. 283, 111979.
Data availability Dikshit, A., Pradhan, B., 2021. Why interpretable machine learning algorithms should be
used in drought forecasting? Natural Hazards Alerts, NSF Convergence Workshop,
24–28 May 2021
The data can be made available from the corresponding author. Francois, C., 2016. Keras, GitHub repository. https://github.com/keras-team/keras.
Fung, K., Huang, Y., Koo, C., Soh, Y., 2020. Drought forecasting: a review of modelling ap-
CRediT authorship contribution statement proaches 2007–2017. J. Water Clim. Chang. 11, 771–799.
Hao, Z., Singh, V.P., Xia, Y., 2018. Seasonal drought prediction: advances, challenges, and
future prospects. Rev. Geophys. 56, 108–141.
Abhirup Dikshit: Conceptualization, Methodology, Modelling, Hayes, M., Svoboda, M., Wall, N., Widhalm, M., 2011. The Lincoln declaration on drought
Writing original draft. Biswajeet Pradhan: Conceptualization, Valida- indices: universal meteorological drought index recommended. Bull. Am. Meteorol.
tion, Visualization, Review & Editing, Supervision, Resource allocation, Soc. 92, 485–488.
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9,
Funding. All authors have read and agreed to the published version of
1735–1780.
the manuscript. Khan, N., Sachindra, D., Shahid, S., Ahmed, K., Shiru, M.S., Nawaz, N., 2020. Prediction of
droughts over Pakistan using machine learning algorithms. Adv. Water Resour. 139,
Declaration of competing interest 103562.
Kirono, D.G., Chiew, F.H., Kent, D.M., 2010. Identification of best predictors for forecasting
seasonal rainfall and runoff in Australia. Hydrol. Process. 24, 1237–1247.
The authors declare that they have no known competing financial Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep
interests or personal relationships that could have appeared to influ- convolutional neural networks. Adv. Neural Inf. Proces. Syst. 25, 1097–1105.
ence the work reported in this paper. Leblanc, M., Tweed, S., Van Dijk, A., Timbal, B., 2012. A review of historic and future hydro-
logical changes in the Murray-Darling Basin. Glob. Planet. Chang. 80, 226–246.
Lloyd-Hughes, B., 2014. The impracticality of a universal drought definition. Theor. Appl.
Acknowledgements Climatol. 117, 607–611.
Lundberg, S., Lee, S.-I., 2017. A Unified Approach to Interpreting Model Predictions arXiv
This research was supported by the Centre for Advanced Modelling preprint arXiv:1705.07874.
Matin, S.S., Pradhan, B., 2021. Earthquake-induced building-damage mapping using ex-
and Geospatial Information Systems, Faculty of Engineering and plainable AI (XAI). Sensors 21, 4489.
Information Technology, University of Technology Sydney. The first McGowan, H.A., Marx, S.K., Denholm, J., Soderholm, J., Kamber, B.S., 2009. Reconstructing
author was supported by the IRTP scholarship funded by the annual inflows to the headwater catchments of the Murray River, Australia, using the
Pacific decadal oscillation. Geophys. Res. Lett. 36.
Department of Education and Training, Govt. of Australia.
BostonMcKee, T.B., Doesken, N.J., Kleist, J., 1993. The relationship of drought frequency
and duration to time scales. Proceedings of the 8th Conference on Applied Climatol-
References ogy, pp. 179–183.
Meteorology, B.o, 2020. Special Climate Statement 70 Update—Drought Conditions in
Abbot, J., Marohasy, J., 2014. Input selection and optimisation for monthly rainfall fore- Australia and Impact on Water Resources in the Murray–Darling Basin, p. 18.
casting in Queensland, Australia, using artificial neural networks. Atmos. Res. 138, Mishra, A.K., Singh, V.P., 2010. A review of drought concepts. J. Hydrol. 391, 202–216.
166–178. Mishra, A.K., Singh, V.P., 2011. Drought modeling–a review. J. Hydrol. 403, 157–175.
Anshuka, A., van Ogtrop, F.F., Vervoort, R.W., 2019. Drought forecasting through statistical Molnar, C., 2020. Interpretable Machine Learning. Lulu Press, Morrisville, NC, USA.
models using standardised precipitation index: a systematic review and meta- Morid, S., Smakhtin, V., Bagherzadeh, K., 2007. Drought forecasting using artificial neural
regression analysis. Nat. Hazards 97, 955–977. networks and time series of drought indices. Int. J. Climatol. 27, 2103–2111.
Ashok, K., Guan, Z., Yamagata, T., 2003. Influence of the Indian Ocean dipole on the Nguyen, H., Wheeler, M.C., Hendon, H.H., Lim, E.-P., Otkin, J.A., 2021. The 2019 flash
Australian winter rainfall. Geophys. Res. Lett. 30 (15). droughts in subtropical eastern Australia and their association with large-scale cli-
Barua, S., Ng, A., Perera, B., 2012. Artificial neural network–based drought forecasting mate drivers. Weather Clim. Extrem. 32, 100321.
using a nonlinear aggregated drought index. J. Hydrol. Eng. 17, 1408–1413. Nicholls, N., 2004. The changing nature of Australian droughts. Clim. Chang. 63, 323–336.

11
A. Dikshit and B. Pradhan Science of the Total Environment 801 (2021) 149797

Nolan, R.H., Boer, M.M., Collins, L., Resco de Dios, V., Clarke, H., Jenkins, M., Kenny, B., human causes and implications for water resources, ecosystems, economy, and soci-
Bradstock, R.A., 2020. Causes and consequences of eastern Australia's 2019–20 season ety. Water Resour. Res. 49, 1040–1057.
of mega-fires. Glob. Chang. Biol. 26, 1039–1041. Verdon-Kidd, D.C., Kiem, A.S., 2009. Nature and causes of protracted droughts in south-
Otkin, J.A., Svoboda, M., Hunt, E.D., Ford, T.W., Anderson, M.C., Hain, C., Basara, J.B., 2018. east Australia: comparison between the federation, WWII, and big dry droughts.
Flash droughts: a review and assessment of the challenges imposed by rapid-onset Geophys. Res. Lett. 36.
droughts in the United States. Bull. Am. Meteorol. Soc. 99, 911–919. Vicente-Serrano, S.M., Beguería, S., López-Moreno, J.I., 2010. A multiscalar drought index
Pendergrass, A.G., Meehl, G.A., Pulwarty, R., Hobbins, M., Hoell, A., AghaKouchak, A., sensitive to global warming: the standardized precipitation evapotranspiration
Bonfils, C.J., Gallant, A.J., Hoerling, M., Hoffmann, D., 2020. Flash droughts present a index. J. Clim. 23, 1696–1718.
new challenge for subseasonal-to-seasonal prediction. Nat. Clim. Chang. 10, 191–199. Vicente-Serrano, S.M., Quiring, S.M., Pena-Gallardo, M., Yuan, S., Dominguez-Castro, F.,
Poornima, S., Pushpalatha, M., 2019. Drought prediction based on SPI and SPEI with vary- 2020. A review of environmental droughts: increased risk under global warming?
ing timescales using LSTM recurrent neural network. Soft. Comput. 23, 8399–8412. Earth Sci. Rev. 201, 102953.
Rudin, C., 2019. Stop explaining black box machine learning models for high stakes deci- Wang, X., Shen, H., Zhang, W., Cao, J., Qi, Y., Chen, G., Li, X., 2015. Spatial and temporal
sions and use interpretable models instead. Nat. Mach. Intell. 1, 206–215. characteristics of droughts in the Northeast China transect. Nat. Hazards 76, 601–614.
Shapley, L.S., 1953. A value for n-person games. Contributions to the Theory of Games. 2, West, H., Quinn, N., Horswell, M., 2019. Remote sensing for drought monitoring & impact
pp. 307–317. assessment: progress, past challenges and future opportunities. Remote Sens. Envi-
Smart, R., 2016. User Guide for Land Use of Australia 2010–11. ABARES, Canberra. ron. 232, 111291.
Ummenhofer, C.C., England, M.H., McIntosh, P.C., Meyers, G.A., Pook, M.J., Risbey, J.S., Yevjevich, V.M., 1967. An objective approach to definitions and investigations of conti-
Gupta, A.S., Taschetto, A.S., 2009. What causes southeast Australia's worst droughts? nental hydrologic drought. Hydrology Paper No. 23. Colorado State University, Fort
Geophys. Res. Lett. 36. Collins, Colorado.
Van Dijk, A.I., Beck, H.E., Crosbie, R.S., de Jeu, R.A., Liu, Y.Y., Podger, G.M., Timbal, B., Viney,
N.R., 2013. The millennium drought in southeast Australia (2001–2009): natural and

12