Prediction droughts using machine learning algorithm

Prediction of droughts over Pakistan using
machine learning algorithms
Presented By
KARTIK JADAV(19AG62R19)
Agricultural and Food Engineering Department, Indian Institute
of Technology Kharagpur, India
Reference : Najeebullah Khan, D.A. Sachindra , Shamsuddin Shahid , Kamal Ahmed a ,
Mohammed Sanusi Shiru , Nadeem Nawaz “Prediction of droughts over Pakistan using
machine learning algorithms ”, Advances in Water Resources(2020) Volume 139, May
2020, 103562

Motivation
 Why study had been done?
 The main objective of this study was to develop ML-based models for
predicting moderate, severe and extreme droughts over Pakistan based on
the SPEI.
 Context of the study
 In this study, the drought prediction models were developed using ANN,
SVM and KNN.
 Contribution
 As droughts have a direct relationship with the availability of water, their
changing characteristics due to changing climate will have a pro- found
impact on water stress and food security.

REVIEW OF LITERATURE
Author Year Findings
Ganguli and
Reddy
2014 SVM is able to learn from a much smaller data set and
is capable of handling a large number of variables. SVM
may overcome certain limitations of ANN such as
trapping at local minima and overfitting, to some extent
in predicting droughts
Yang et al 2015 Among the ML techniques, ANN and SVM can be
regarded as the most widely used techniques for
developing drought prediction models

Study area and datasets
• Pakistan located in South
Asia covers an area of
796,095 km2.
• Pakistan experiences two
monsoon precipitation
seasons; IM (July–Sep),
WD (Dec–Mar).
Fig. 1. IM refers to the regions which receive
precipitation mainly during the Indian monsoon and
WD refers to the regions which receive
precipitation mainly during the western
disturbance.

• It experiences four seasons (based on temperature) cool and dry winter
(Dec–Feb), hot and dry spring (Mar–May), hot and humid summer (Jun–Aug)
and dry autumn (Sep–Nov)
• During these two precipitation seasons, two important cropping seasons
called Kharif (May-Oct) observed during the Indian monsoon period
• The Rabi (Nov-Apr) is experienced during the western disturbances
• These cropping seasons play an important role in agriculture and the
economy of Pakistan

Gridded precipitation and
temperature datasets
• This study uses the Princeton global forcing (PGF) gridded temperature
and precipitation datasets for estimating the SPEI (NCEP/NCAR datasets)
• The PGF gridded temperature and precipitation datasets have a resolution
of 0.25 ×0.25 in the latitudinal and longitudinal directions and hence the
data are available at 1437 grid points over Pakistan.
• These data were used in the calculation of SPEI at each of the 1437 grid
points distributed over Pakistan, Among which 531 grid points lie in the
region of WD and 906 grid points lie in the IM dominated region

Method
1–6-month SPEIs were calculated for each of the grid points
Based on 1–6-month SPEIs droughts were categorized
Different large-scale atmospheric variables (probable predictors) that are
influential on droughts were identified based on past literature
Principle Component Analysis (PCA) was used to generate the Principal
Components (PCs) from the data
The difference between the PCs for February and April(Kharif) and difference
between the PCs for October and August(Rabi) were computed.
The PCs selected using SVM-RFE they were used in the development of
drought prediction models
the correlations between the PCs and 1–6-month SPEI were separately
calculated
ML-algorithms were used to develop drought prediction

PCA
Why?
• Overfitting (Due to too much features and attribute we gave
during training phase so model is confused and problem of
overfitting occur)
• PCA is Linear projection method to reduce number of
parameters(Attributes and features).
1. Standardize the data
2. Calculate the covariance matrix
3. Find the eigenvalues and eigenvectors of covariance matrix
4. Plot the eigenvectors/principal components over scaled data

SVM
• SVM is a widely used ML algorithm that can be used in developing
classification and regression models.
• However, Sachindra et al. (2018) showed that the polynomial kernel
performs better with SVM. Hence, the polynomial kernel was used in this
study.
• The polynomial kernel is defined as
. 𝐾 (𝑥 𝑖 , 𝑥 𝑗 )= (𝑥 𝑖 . 𝑥 𝑗 + 𝑐 ) 𝑑
Where, 𝑥 𝑖 and 𝑥 𝑗 are the predictor and predictand data, d is the degree of
the polynomial, and c is a constant that allows a trade-off between the
influence of the higher and lower order terms.

ANN
• .To reduce the chances of overfitting, a technique known as Bayesian
regularization was developed ( Burden and Winkler, 2008 ; MacKay,
1992 ; Ticknor, 2013 ).
𝐹 = 𝛽E D + 𝛼E w
• In the Bayesian network, the weights are considered random variables
and thus their density function is expressed following the Bayes’ rules
𝑃 ( 𝑤 |𝐷, 𝛼, 𝛽, M) =
𝑃 (D |w , 𝛽, M) P(w |𝛼, M)
𝑃 (D |𝛼, 𝛽, M)
Where w is the vector of network weights, D
represents the predictor and the predictand data
vectors (x, y), and M is the neural network model
being used, 𝛼 and 𝛽 are Regularization

K-Nearest Neighbour
• The KNN for classification and regression is the simplest non-parametric ML
techniques
• KNN algorithm finds the k Nearest Neighbours from calibration data using a
distance measure such as the Euclidean distance
𝑑 𝑗𝑜 = 𝑖=1
݊ (𝑥 𝑖𝑗 − 𝑥 𝑖𝑜 )2 𝑡 = 1 , 2 , 3 , …., ݊
• For calculating KNN simulated values ( Zr ) Zk refers to the neighbouring data
and f k ( dj ) is the kernel function
𝑍 𝑟 = ݇ =1
𝐾 ݂݇ (𝑑𝑗 ×𝑍݇ ).

Performance assessment
1) The coefficient of determination (R2 )
varies between 0 (no agreement) and 1
(perfect agreement)
2) The normalized root mean squared error
(NRMSE) varies between 0 (perfect
agreement) and +∞(no agreement)
3) Percentage of bias (Pbias) varies between
- ∞(underestimation) and +∞(over-
estimation) and a value of 0 refers to a
perfect agreement
4) lThe modified index of agreement ( md )
varies between 0 (no agreement) and 1
(perfect agreement)
N is the number of testing samples. Oi are
the ith observation, Si are the i th simulated
value, sd is the standard deviation of the
observations.

Results and Discussion
Fig. 2. Climate domain used in the present study, 861 NCEP/NCAR grid
points spread across the climate domain are depicted in grey circles.

Fig. 3. The spatial pattern of the correlation coefficient of the RH, atm temperature and wind
component (925hPa) during the Rabi and Kharif periods. The correlation is calculated with
6-month SPEI in April for Rabi and in October for Kharif. The dots represent the significant
correlation at 95% confidence level with the SPEI.

Fig. 4. Boxplots showing correlations of the selected atmospheric variables at four
pressure levels i.e. 925, 850, 700 and 500 hPa (except SLP) with the SPEI of Pakistan
Red boxplots (Rabi season) Blue (Kharif season)

Fig. 5. Performance of different machine learning-based models in
predicting moderate droughts in the Rabi season.

Fig. 6. Performance of different machine learning-based models
in predicting moderate droughts in the Kharif season.

Fig.7. Boxplots of the spatial correlation and NRMSE for the Rabi and Kharif
seasons for the month of April shown in red and blue colours respectively.

Conclusions
• In this study, it was seen that KNN-based drought models display limited
performance comparison to other drought models
• SVM-based models were able to better capture the temporal and spatial
characteristics of droughts
• It was found that in the Rabi season SPEI is positively correlated with RH
over the Mediterranean Sea and the region north of the Caspian Sea,
whereas in the Kharif season SPEI is positively correlated with the humid
region over the south-eastern part of the Bay of Bengal and Caspian Seas.

• Rabi season SPEI also displayed statistically significant positive
correlations with atmospheric temperature over the region north of the
Mediterranean Sea.
• Also, wind speeds over the Indian Ocean and the Arabian Sea during
the Kharif season were statistically significantly correlated with SPEI
.
• This indicated that RH, temperature and wind speed are indicators of
droughts over Pakistan.

Limitations of this study
• The present study used only one reanalysis dataset (NCEP/NCAR) to provide
inputs to the drought prediction models.
• In order to study the uncertainties which originate from the inputs, different
reanalysis datasets should be used in developing drought prediction models.
• However, this was not performed as the other major reanalysis datasets
either did not have data for some of the specific atmospheric variables used
in this study or they did not have data corresponding to the period 1948–
2016.

Future scope
• Sachindra and Kanae (2019) proved that Parallel Multi-Population Genetic
Programming (PMPGP)-based statistical downscaling models show better
generalization skills and higher resistance to redundant information in
inputs compared to those based on traditional GP.
• Thus, other ML algorithms such as Extreme Learning Machine (ELM),
Genetic Programming (GP), PMPGP, and Random Forest (RF) should also
be tested in future for developing drought prediction models over Pakistan.

References
• Najeebullah Khan, D.A. Sachindra , Shamsuddin Shahid ,
Kamal Ahmed a , Mohammed Sanusi Shiru , Nadeem Nawaz
“Prediction of droughts over Pakistan using machine learning
algorithms ”, Advances in Water Resources(2020) Volume
139, May 2020, 103562
• Ganguli, P. , Reddy, M.J. , 2014. Ensemble prediction of
regional droughts using climate inputs and the SVM–copula
approach. Hydrol. Process. 28, 4989–5009

Prediction droughts using machine learning algorithm

Prediction droughts using machine learning algorithm

More Related Content

What's hot (20)

Similar to Prediction droughts using machine learning algorithm (20)

Recently uploaded (20)

Prediction droughts using machine learning algorithm