Incorporation of Climate Change Effect on Indian Hydroclimatology
Dr. Rajib Maity
Assistant professor, IIT Bombay, India
Current Position: Honorary Visiting ScholarSchool of Chemistry, Physics and Earth Sciences
Flinders University, Adelaide, Australia
June 18,2008 2
Introduction
Overview of climate change impact
Indian hydroclimatology: Continental scale
Indian Summer Monsoon Rainfall (ISMR)
Large-scale coupled circulation pattern: ENSO and EQUINOO
Prediction of ISMR using ENSO and EQUINOO
Indian hydroclimatology: Basin scale
ENSO, EQUINOO and Basin-Scale Streamflow
Traditional modeling approaches vs. hydroclimatological approach
Conclusions
Outline
June 18,2008 3
Climate Change Overview
Web Source: http://geoscape.nrcan.gc.ca/h2o/bow/climate_e.php
How to model hydrologic time series under climate change scenario?
June 18,2008 4
AtmosphericScience
Hydrology and Water Resources
Overview
Hydroclimatic Teleconnection: The significant association between hydrologic events and large-scale atmospheric circulation patterns, which are widely separated (planetary scale) across the globe, is referred as ‘hydroclimatic teleconnection’.
June 18,2008 5
Natural variability of hydrologic variables and hydroclimatic teleconnection
Volu
me
(M A
cre-
Feet
)
Year
It is recently being established that temporal structure of hydrologic time series is significantly forced by large-scale atmospheric circulation patterns throughhydroclimatic teleconnection (Jain and Lall, 2001)
June 18,2008 6
Major Issues/Challenges
• Dynamic nature of cause-effect relationship between hydrologic time series and large-scale circulation is to be captured.
• Uncertainty associated with the prediction to be addressed
• Consideration of nonstationarity in time series, expected under climate change scenario.
• ‘Large-Scale’ to ‘Basin-Scale’: It is necessary to establish the link.
• Being most essential for Hydrology and Water Resources responses of ‘basin-scale’ hydrologic variables to ‘large-scale’ atmospheric circulation will be investigated
• Use of such relationship with a goal to improve the prediction performance of hydrologic time series.
June 18,2008 7
Source: NOAA
El Niño-Southern Oscillation (ENSO) is the coupled Ocean-atmosphere mode of tropical Pacific Ocean (Cane, 1992)
Attempts were made to forecast hydrologic variables, like rainfall, streamflow, etc., using ENSO information all over the world (Dracup and Kahya, 1994; Eltahir, 1996; Jain and Lall, 2001).
El Niño-Southern Oscillation(ENSO)
June 18,2008 8
El Niño condition La Niña condition
SST(°C)
AnomalousSST (°C)
Source: NOAA
El Niño-Southern Oscillation (ENSO)
AnomalousPressure
(mb)
June 18,2008 9
Source: NOAA
El Niño-Southern Oscillation(ENSO)
of 38
June 18,2008 10
Source: NOAA
El Niño conditionNormal conditionLa Niña condition
Change in Atmospheric Circulation Pattern
of 38
June 18,2008 11
Impact of El Niño
Rainfall defienciesAfrica, the Indian Subcontinent, northern China, Australia, northern South America and the Caribbean
Low river dischargeNile (Egypt), Senegal (Senegal), Orange (South Africa), Krishna (India), Murray-Darling (Australia) and Amazon (Brazil) river system
Surplus rainfallWestern Europe, eastern Africa, Western Cape of South Africa, Gulf regions of the united states and northern Mexico and central Chile.
Sources: NOAAAllan et al., 1996
Additional Slide #
June 18,2008 12
Impact of La Niña
Increased RainfallAfrica, the Indian Subcontinent, northern China, Australia, northeastern South America
High River DischargeNile, Senegal, Orange, Krishna, Murray-Darling and Amazon
Enhanced Monsoonal/ Tropical-linked summer rainfallSouthern Africa and India, and prominent cloud band influences lead to rainfall surplus regions over southern Australia and southern Africa
Sources: NOAAAllan et al., 1996
June 18,2008 13
Niño Regions
Niño 3.4 sea surface temperature anomaly (Niño 3.4 SSTA) is best correlated with Indian summer monsoon rainfall
June 18,2008 14
Equatorial Indian Ocean Oscillation (EQUINOO)
Atmospheric component of Indian Ocean Dipole (IOD) Mode is known as EQUINOO.
EQUINOO is defined as the oscillation between two opposite states of convection over EEIO and WEIO.
Equatorial zonal wind index (EQWIN) is considered as an measure of this which is the negative of the anomaly of zonal wind at Eq. I O Reg.
EEIO = 90° – 110° E, 10° S – 0°WEIO = 50° – 70° E, 10° S – 10° NEq I O Reg = 60° – 90° E, 2.5° S – 2.5° N
June 18,2008 15
Bayesian Dynamic Linear Model (BDLM)
Dynamic nature of the model helps to capture the dynamic relationship between climate information and hydrologic variables
Allowance incorporation of exogenous inputs
Relaxation of stationarity assumption
Allowance for on-line external intervention
Probabilistic Forecast
Important Features
June 18,2008 16
Model Calibration
Calibration Period: 1959 – 1985
Performance Statistics
• Correlation coefficient (CC) between observed and predicted rainfall anomaly
• CC between observed and predicted rainfall • Log-likelihood• Mean Square Error (MSE)
Testing Period: 1986 – 2003
Additional Slide #
June 18,2008 17
Model Performance
June 18,2008 18
Model Performance
Correlation coefficient between observed and predicted rainfall during monsoon months = 0.82
June 18,2008 19
Model Performance
June 18,2008 20
Comparative Performance
a) Only ENSO index,
b) Only EQUINOO index and
c) Both ENSO and EQUINOO indices
June 18,2008 21
Semiparametric approach using Copula
A copula is a function that joins or couples multiple distribution functions to their one-dimensional marginal distribution functions.Let X and Y be a pair of ‘Random Variables’ with cdf as FX(x) and GY(y). Also let their joint cdf as HX,Y(x,y).
( )xFx
( )yGy
( )0,1
( )1,0
( )0,0
( )yxH yx ,,
0 1( )yx,
June 18,2008 22
Observations (1958 – 1985) Observations (1986 – 2003)
Simulation with different copulas
June 18,2008 23
Comparison between observed and predicted rainfall anomaly
June 18,2008 24
Spatial Variation
June 18,2008 25
At monthly time-scale, the hydroclimatic teleconnection between basin-scale streamflow and large-scale atmospheric circulation pattern is more complex.
Genetic Algorithm based Evolutionary Optimizer is opted to obtained optimized Artificial Neural Network (ANN), which is used to captured the complex relationship
Sub-basin Map of Mahanadi River
Monthly streamflow are measured at Basantpur
June 18,2008 26
Stepsa) Parameters initialization: All the parameters like population size (N), number of
maximum generations, probability of crossover (Pc), and probability of mutation (Pm) are set to specific values. N = 50, Pc = 0.2, Pm = 0.04
b) Generation of initial population: At initial generation, the evolutionary optimizer randomly creates N networks for the initial population
c) Training of the network and fitness evaluation: Networks are trained by back propagation algorithm and their fitness values are determined according to the goals to be achieved. These goals are (i) maximum number of neurons used in the network (Model Parsimony), (ii) the mean square deviation and (iii) maximum square deviation for the training set. Weighted average (expressed as percentage) of the goal parameters is used as the fitness of the individual network. Based on the fitness values, ranking of the network is done
Genetic Algorithm based Evolutionary Optimizer
June 18,2008 27
d) Evolution of new generation:
i) Two “parent” networks will be chosen out of the old generation. The selection algorithm will choose networks with a high fitness by a higher probability.
ii) Two “children” networks will be created from the two “parent” networks. Using the cross over probability (Pc), the two “parent” networks will be crossed over, i.e., they will swap a portion of the network with each other.
iii) The “children” will be mutated with a mutation probability (Pm)
iv) A few elitist members of the population in current generation are also carried to the next generation in addition to the “children”. The selection continues until the new generation also has N members. After completion, the new generation will be evaluated.
e) Checking for Termination Criteria: Evolution is stopped, if the target goals are achieved at least for one network in the population or maximum number of generations is reached. Otherwise, generation counter is increased by one and steps (c) and (d) are repeated.
Genetic Algorithm based Evolutionary Optimizer
June 18,2008 28
Model Performance for Testing Period (1992 – 2003)
June 18,2008 29
Predicted Category of StreamflowObserved Category of Streamflow Low Normal High
Low 12 6 3
Normal 3 20 4
High 0 3 9
( ) ( )randrand CFNCFCFHSS −−=Heidke skill score, = 0.525(Wilks, 2006)
Model Performance for Testing Period (1992 – 2003)
June 18,2008 30
Considering only streamflow information of previous month(s)
Considering only large-scale atmospheric circulation
information
Considering both streamflow (if necessary) and large-scale
atmospheric circulation information
Inputs* Best Network
Model Perfor-mance#
Inputs* Best Network
Model Perfor-mance#
Inputs* Best Network
Model Perfor-mance#
Jun May SF 1-2-2-1-0.030833.01500.3
Mar EN Mar EQ 2-4-5-3-1
0.5421209.41926.4
Mar EN Mar EQ 2-4-5-3-1
0.5421209.41926.4
Jul Jun SF 1-2-2-20.7314312.85025.6
Jun EN Jun EQ 2-3-2-1
0.2985643.66059.4
Jun SF Jun EN Jun EQ
3-3-5-10.8133570.94425.2
Aug Jul SF 1-2-3-1-10.4122929.23575.7
Jul EN Jul EQ 2-2-1
0.5112721.53225.6
Jul SF Jul EN Jul EQ
3-6-5-10.8361486.02008.5
Sep Jul SF Aug SF 2-2-3-1-1
0.3212331.93260.2
Aug EN Aug EQ 2-3-1-2-1
0.4021877.12719.7
Jul SF Aug SF Aug EN Aug EQ
4-3-2-10.4932101.02635.6
Oct Sep SF 1-2-10.838611.6761.4
Sep EN Sep EQ 2-4-3-1
0.372898.8
1228.1
Sep SF Sep EN Sep EQ
3-5-10.765742.2926.5
Month of Strea-
mflow
Performance for three different cases
June 18,2008 31
Conclusions
Bayesian dynamic linear model (BDLM) is introduced to the field of hydroclimatology, which is able to capture the time varying dynamic relationship and quantify the uncertainty associated with the predicted values.
A Copula-based method is able to capture the scale-free dependence pattern and provides a nonparametric way to predict the response variable using the information from causal force.
These methods, being general, can be applied to any other similar analysis. Copula-based method is applicable irrespective of the nature of dependence, type of relationship and distributional properties of data sets.
June 18,2008 32
Conclusions
Mahanadi River basin is selected to study the basin-scale hydroclimatic teleconnection. Hydroclimatic association at seasonal scale between monsoonal inflow into Hirakud reservoir on Mahanadi River and large-scale circulation phenomena is established.
At monthly scale, variation of streamflow at Basantpur site on Mahanadi River is investigated. The information of streamflow from previous month(s) alone may not be sufficient.
Consideration of both the information of previous streamflow and the two large-scale atmospheric circulations is important for basin-scale monthly streamflow prediction improves the prediction performance .
June 18,2008 33