Vous êtes sur la page 1sur 9

ANN-Based Bias Correction Algorithm

for Precipitation in the Yarra River Basin,


Australia

P. Saravanan1 , C. Sivapragasam1(B) , M. Nitin2 , S. Balamurali3 , R.K. Ragul1 ,


S. Sundar Prakash1 , G. Selva Ganesan1 , and V. Vel Murugan1
1
Centre for Water Technology, Kalasalingam University, Srivilliputtur, India
p.saravanan.me.1986@gmail.com, sivapragasam25@gmail.com
2
College of Engineering and Science, Victoria University, Melbourne, Australia
Nitin.Muttil@vu.edu.au
3
Department of Computer Applications, Kalasalingam University,
Srivilliputtur, India
sbmurali@rediffmail.com

Abstract. Regional Climate Models (RCM) applied to simulate future


climate parameters such as precipitation and temperature are reported
to suffer from bias. Bias correction is necessary for using such data for
climate change impact studies. In this study, a new ANN based bias
correction algorithm is suggested and is compared with other three
conventional methods, namely linear scaling, local intensity and power
transformation. The proposed method outperforms conventional meth-
ods with mean, standard deviation and the RMSE of bias corrected time
series more closely matches with that of the observed precipitation.

Keywords: Artificial Neural Network · Bias correction · Precipitation ·


Simulation · Climate change projection

1 Introduction

Precipitation is an important meteorological parameter which influences the


flood and drought situations of any country. Climate change is believed to have
perceptible effect on the occurrence of precipitation and its distribution and its
quantification is a direct concern of water resources managers. Conventionally,
future projected precipitation is simulated by Regional Climate Models (RCM)
for climate change scenarios suggested by Intergovernmental Panel on Climate
Change (IPCC). Precipitations simulated by RCMs are found to contain sys-
tematic error (also called as bias) due to inaccurate parameterization of the
climatic process during the model development [5]. Bias means, deviation of
statistics like the mean, variance, covariance of the model from the correspond-
ing observed value [9,13]. Such simulated information need to be bias corrected
before it can be used for any hydrological studies [1,4,7,14,18].


c Springer International Publishing AG 2017
S. Arumugam et al. (Eds.): ICTCSDM 2016, LNCS 10398, pp. 362–370, 2017.
DOI: 10.1007/978-3-319-64419-6 47
ANN-Based Bias Correction Algorithm for Precipitation 363

Considerable research has already been done for bias correction. Some
authors compared existing methods to find the best method for their own
model [20], whereas other authors have proposed new bias correction methods
[16,17,19,21,28]. Teutschbein and Seibert in [27] reviewed different bias correc-
tion methods such as quantile mapping, power transformation, local intensity
and linear scale by comparing their performance both in terms of deviation from
the observed precipitation as well as end use application of stream flow sim-
ulation. The performance of quantile mapping and power function are found
to be the most robust. Some studies such as Tschoke et al., in [26] focused on
developing new methodologies for error reduction during dry periods. Methods
adopted in bias correction are seen to vary from very simple methods to advanced
methods. While simple methods are found to perform poorly in summer season,
the advanced methods offer difficulties in terms of long data length required to
calibrate [3].
In a recently reported work, Um et al., in [28] proposed a hybrid bias cor-
rection method and compared that with other two conventional methods viz.,
linear scaling and quantile mapping. While linear scaling yielded the best result
for estimating annual average precipitation, the hybrid method was reported to
be optimal for predicting the variation in annual precipitation.
The bias corrections also find many applications other than for precipitations.
For instance, Ahmed et al. [1] bias corrected a data set of daily maximum and
minimum temperature for direct use of climate change impact studies for the
future period of 2046–2065. Macias et al. [22] simulated the sea surface tempera-
ture using different ocean model and compared the result with satellite observed
data and identified the bias of different models. They also applied simple bias
correction to atmospheric variables of the model, to know the importance of
each variable and found that wind velocity is the most important variable to
bias correct.
Although the conventional bias correction methods are most popularly
adopted, of late, researchers started applying black box methods such as Artifi-
cial Neural Network (ANN). Sanaz Moghim, in [24] have applied ANN for bias
correction of Precipitation and Temperature. Also Chitra and Santhosh in [6]
used ANN to downscale the simulated data. They applied Delta Method for bias
correction.
To the author’s knowledge, the application of ANN for bias correction is
still in its inception with very few reported works. In this study we proposed
a new robust bias correction algorithm which while reducing the Root Mean
Squared Error(RMSE) between observed and simulated data, also tries to map
the mean and standard deviation of the observed precipitation. The results are
compared to the conventional methods of local intensity, linear scaling, and
power transformation.

2 Study Area and Data Set


The middle segment of Yarra River Catchment is used for this study (Fig. 1).
The water resources management is complex in this catchment due to the need
364 P. Saravanan et al.

of downstream supply for Melbourne as well as environmental flow provisions [2].


This catchment is the fourth highest productive in Victoria, even though its area
is comparatively less than other catchments [23].
Observed daily precipitation data from 1980 to 2012 are available for the
Yarra River Catchment, and are used in this study. Simulated data are obtained
from the Climate change Data for SWAT (CMIP3) [8] database for the year 1961
to 2000 using CCCMA CGCM3.1 model. For the bias correction analysis, only
the overlapping time period (1980 to 2000) is used for analysis.
For the analysis we considered two stations located within the middle segment
of Yarra River Catchment, which are Toolangi (S -37.57, E 145.5) and Black Spur
(S-37.59, E 145.62). The locations of these stations can be seen in Fig. 1. The
observed daily rainfall data of Toolangi and Black Spur and the nearest available
simulated data of the catchment is taken for this analysis.

Fig. 1. Location of study area

3 Methods
In this section, the methodologies of four bias correction methods used in this
study are explained. Three of them are conventional methods. The methodology
of the three conventional methods is explained based on [10,27].

3.1 Linear Scaling Method


This method is a mean based method which considerably reduces the deviation in
the mean of observed and simulated data. The observed precipitation is corrected
by a factor which is the ratio of long term monthly mean of observed and raw
simulated precipitations
μ(Pobser )(m)
Pcorrect(m)(d) = Praw(m)(d) (1)
μ(Praw )(m)
Here, Pcorrect(m)(d) and Praw(m)(d) , are the corrected and raw simulated pre-
cipitation for dth day of mth month respectively.
ANN-Based Bias Correction Algorithm for Precipitation 365

μ() , represents mean operator. For example μ(Praw )(m) is the mean value of raw
precipitation for the given month.

3.2 Local Intensity Method

This method is an improved version of linear scaling which not only corrects the
monthly mean but also the wet day frequency and wet day intensity. The bias
correction is done as follows:
Initially a threshold Pthreshold(m) of simulated data is calculated such that
the number of days in a specified month which is more than the threshold pre-
cipitation equals the number of wet days (day of non zero precipitation) in the
observed data.
Then the corrected simulated data is calculated as follows.


0 if Praw(m)(d) < Pthresold(m)
Pcorrect(m)(d) =
(Praw(m)(d) ).(S) Otherwise

Here ‘S’ is a scaling factor considering only wet day of observed and wet day
of corrected simulated data and is calculated as follows,
μ((Pobser )(m)(d) | (Pobser )(m)(d)>0))
S= (2)
μ((Praw )(m)(d) | (Praw )(m)(d)>(Pthreshold )(m))

3.3 Power Transformation

The power transformation method can correct the standard deviation which
is difficult to be ensured through linear scaling and local intensity methods.
However, this method uses the bias corrected data by local intensity as input.
Initially a parameter b is calculated for each month m by using coefficient of
variation (CV) of data corrected by local intensity and CV of observed data:

f (bm ) = CVPobser(m) − (CVPcorrect(LOI)(m) )bm (3)


Here, bm is an exponent of mth month.
After the determination of ‘bm . ‘S’ is calculated as follows:
μ(Pobser )(m)
S= (4)
μ(Pcorrect(LOI)(m) )bm
The bias corrected value obtained by power transformation is calculated as
follows:
Pcorrect(m)(d) = (S).(Pcorrect(LOI)(m)(d) )bm (5)
where Pcorrect(m)(d) andPcorrect(LOI) are the corrected data by Power Transfor-
mation and Local Intensity respectively
366 P. Saravanan et al.

3.4 ANN-based Bias Correction Algorithm

Fundamental Concepts of ANN: Artificial Neural Network (ANN) is a net-


work of many simple elements called neurons, each having a small amount of
local memory. The neurons are connected through communication channels or
connections which carry numeric data encoded by various means. Each neuron
operates only when it receives data through the communication channels. The
architecture of ANN is motivated by the structure of the human brain and nerve
cells. Historically, much of the inspiration to build ANNs came from the desire
to produce artificial systems capable of sophisticated computations similar to
those of the human brain.
Neural network modeling is based on the use of architecture and learning
paradigms which allow the extraction of statistical structure present in the data
set. This ‘connectionist’ philosophy is based on the idea that with a general
architecture, often biologically motivated and with no prior information about
the phenomenon of interest, it is possible to ’learn’ the underlying structure
of the data in an implicit form. The acquired information about the data is
‘stored’ at the connections between the elements of the neural architecture. The
architecture is initially not structured, and the learning algorithm is responsible
for the extraction of the regularities present in the data by finding a suitable set
of synapses during the process of observation of the examples. Thus, ANNs solve
problems by self-learning and self-organization. They derive their ‘intelligence’
from the collective behavior of simple computational mechanisms at individual
neurons.
Multilayer feed-forward network with back-propagation learning algorithm
is one of the most popular neural network architectures, which has been deeply
studied and widely used in many fields. The feed-forward architecture allows
connections only in one direction, that is, there is no back-coupling between
neurons, and the neurons are arranged in layers, starting from an input layer and
ending at the final output layer with one or more hidden layers. The information
passes from the input to the output side. Each layer is made-up several neurons,
and the layers are interconnected by a set of weights. The neurons in the input
layers receive input directly from the input variables. The neurons in the hidden
and output layers receive input from interconnections. Neurons operate on the
input and transform it to produce an analog output. More details on ANN can
be found in [11,12,15,25].

Model Development: Two different ANN models are employed in this study,
which are discussed below.

Model 1(ANN-M1): Since the aim is to determine the bias corrected precipitation
from the simulated precipitation, in the first model, a direct mapping is done with
simulated precipitation as the input and observed precipitation as the output.
It is desired to minimize the Root Mean Square Error (RMSE) of the corrected
precipitation besides ensuring a closer match with standard deviation and mean
of the observed precipitation. A three layer feed forward network with 11 hidden
ANN-Based Bias Correction Algorithm for Precipitation 367

nodes gives the best output. Out of a total of 253 data of monthly precipitation,
108 is used for training, 24 for testing and remaining for verification.
Model 2(ANN-M2): Instead of directly mapping simulated and observed precip-
itation, it is desired to map the simulated precipitation to the absolute value
of difference between observed and simulated (i.e. error in the simulated and
observed precipitation, |Δe|). The number of training, testing and verification
data is kept same as that used in Model 1. The output from the training has to be
re-corrected to obtain the actual precipitation for which the following algorithm
is used:

Psim + Δe, if Psim (Psim,i ) − (Psim,j ) and θ(Δe) > 50
Pcorrected = (6)
Psim + Δe, Otherwise

where the (Psim,i ) and (Psim,j ) are the ranges adopted in this study is listed in
Table 1;Δe = (Pobs,i ) − (Psim,i );θ(Δe) is the percentage of positive error in the
specified range i.e.(Psim,i ) − (Psim,j )
This algorithm might induce significant error for those ranges of simulated
precipitation for which θ(Δe) is in the neighborhood of 50.

4 Results and Discussions

In this section, the bias corrected simulated precipitation by the methods of lin-
ear scaling, local intensity, power transformation, ANN-M1, ANN-M2 are com-
pared with the corresponding observed precipitation and is tabulated in Table 2.
Model 1 (ANN-M1): From Table 2, it is observed that the ANN-M1 has not only
the lowest RMSE when compared to all the other methods, but also the SD
for both the stations is also very low. Low SD indicates that the prediction has
failed in capturing the variations in the precipitation as seen from Figs. 2 and
3. The conventional methods perform almost equally well for both the stations
in terms of all the performance measure considered. Figures 2,3 indicates that

Fig. 2. Comparison of various methods for Black Spur Station


368 P. Saravanan et al.

Fig. 3. Comparison of various methods for Toolangi Station

Table 1. Applied error sign for various range

Station name (Psim,i ) − (Psimj ) θ(Δe) Error Sign (+/−)


Toolangi 0–30 5.55 +
30–60 11.42 +
60–90 12.50 +
90–120 28.57 +
> 120 76.92 −
Black Spur 0–30 10.52 +
30–60 18.42 +
60–90 19.23 +
90–120 33.33 +
> 120 88.00 −

power transformation method overestimates the peaks to a much higher degree.


This behavior can be reduced if RMSE can be reduced while maintaining the
mean and SD.
Model 2 (ANN-M2): As seen from Table 2, the proposed algorithm increases
the SD while reducing the RMSE. The RMSE is much less when compared to

Table 2. Results of various statistics for different bias correction methods

Station Statistics Observed Simulated Bias Bias Bias Bias Bias


precipita- precipita- corrected corrected corrected by corrected corrected
tion tion by linear by local power trans- by ANN by ANN
scaling intensity formation model 1 model 2
Toolangi Mean 119.14 75.99 125.36 122.82 125.76 129.70 127.41
SD 58.48 44.00 78.44 78.13 70.54 0.82 34.69
RMSE 87.84 99.67 99.46 95.50 59.10 67.76
Black Spur Mean 115.27 85.96 117.41 114.36 110.70 121.63 127.20
SD 62.80 55.97 79.95 76.23 76.14 15.65 37.43
RMSE 89.53 95.94 94.90 95.31 64.28 70.22
ANN-Based Bias Correction Algorithm for Precipitation 369

the conventional methods, and the SD has been considerably increased when
compared to ANN-M1. The effect of this can be clearly seen in Figs. 2 and 3 in
terms of mapping the peaks and other values more closely when compared to
the conventional methods.

5 Conclusions
Based on this study the following conclusions can be drawn:
(a) ANN seems to be a potential tool for bias correction
(b) The proposed algorithm is able to correct the simulated precipitation to map
more accurately with the observed value when compared to the conventional
methods of bias correction.

References
1. Ahmed, K.F., Wang, G., Silander, J., Wilson, A.M., Allen, J.M.: Statistical down-
scaling and bias correction of climate model outputs for climate change impact
assessment in the US northeast. Global Planet. Change 100, 320–332 (2013)
2. Barua, S., Muttil, N., Ng, A.W.M., Perera, B.J.C.: Rainfall trend and its impli-
cations for water resource management within the Yarra River catchment. Aust.
Hydrol. Process. 27(12), 1727–1738 (2013)
3. Berg, P., Feldmann, H., Panitz, H.J.: Bias correction of high resolution regional
climate model data. J. Hydrol. 448, 80–92 (2012)
4. Ceglar, A., Kajfež-Bogataj, L.: Simulation of maize yield in current and changed
climatic conditions: addressing modelling uncertainties and the importance of bias
correction in climate model simulations. Eur. J. Agron. 37(1), 83–95 (2012)
5. Chen, J., Brissette, F.P., Chaumont, D., Braun, M.: Finding appropriate bias cor-
rection methods in downscaling precipitation for hydrologic impact studies over
North America. Water Resour. Res. 49(7), 4187–4205 (2013)
6. Chithra, N.R., Santosh, G.T.: Bias correction of ANN based statistically down-
scaled precipitation data for the Chaliyar river basin. Int. J. Innov. Res. Sci. Eng.
Technol. 2, 6–11 (2013)
7. Christensen, J.H., Boberg, F., Christensen, O.B., Lucas-Picher, P.: On the need
for bias correction of regional climate change projections of temperature and pre-
cipitation. Geophys. Res. Lett. 35(20), 1–6 (2008)
8. Climate change Data for SWAT (CMIP3) Database. http://globalweather.tamu.
edu/cmip (Viewed October 2016)
9. Ehret, U., Zehe, E., Wulfmeyer, V., Warrach-Sagi, K., Liebert, J.: HESS opinions
“Should we apply bias correction to global and regional climate model data?”.
Hydrol. Earth Syst. Sci. 16(9), 3391–3404 (2012)
10. Fang, G., Yang, J., Chen, Y.N., Zammit, C.: Comparing bias correction methods
in downscaling meteorological variables for a hydrologic impact study in an arid
area in China. Hydrol. Earth Syst. Sci. 19(6), 2547–2559 (2015)
11. Govindaraju, R.S.: Artificial neural networks in hydrology. I: preliminary concepts.
J. Hydrol. Eng. 5(2), 115–123 (2000)
12. Govindaraju, R.: Artificial neural networks in hydrology: II, hydrologic applications
(2000)
370 P. Saravanan et al.

13. Haerter, J.O., Hagemann, S., Moseley, C., Piani, C.: Climate model bias correction
and the role of timescales. Hydrol. Earth Syst. Sci. 15(3), 1065–1079 (2011)
14. Hawkins, E., Osborne, T.M., Ho, C.K., Challinor, A.J.: Calibration and bias correc-
tion of climate projections for crop modelling: an idealised case study over Europe.
Agric. For. Meteorol. 170, 19–31 (2013)
15. Haykin, S.S.: A Comprehensive Foundation. Tsinghua University Press, Bejing
(2001)
16. Hoffmann, H., Rath, T.: Meteorologically consistent bias correction of climate time
series for agricultural models. Theor. Appl. Climatol. 110(1–2), 129–141 (2012)
17. Ines, A.V., Hansen, J.W.: Bias correction of daily GCM rainfall for crop simulation
studies. Agric. For. Meteorol. 138(1), 44–53 (2006)
18. Johnson, F., Sharma, A.: What are the impacts of bias correction on future drought
projections? J. Hydrol. 525, 472–485 (2015)
19. Kim, K.B., Kwon, H.H., Han, D.: Bias correction methods for regional climate
model simulations considering the distributional parametric uncertainty underlying
the observations. J. Hydrol. 530, 568–579 (2015)
20. Lafon, T., Dadson, S., Buys, G., Prudhomme, C.: Bias correction of daily precip-
itation simulated by a regional climate model: a comparison of methods. Int. J.
Climatol. 33(6), 1367–1381 (2013)
21. Li, J., Sharma, A., Evans, J., Johnson, F.: Addressing the mischaracterization of
extreme rainfall in regional climate model simulations-a synoptic pattern based
bias correction approach. J. Hydrol. doi:10.1016/j.jhydrol.2016.04.070
22. Macias, D., Garcia-Gorriz, E., Dosio, A., Stips, A., Keuler, K.: Obtaining the
correct sea surface temperature: bias correction of regional climate model data for
the Mediterranean Sea. Clim. Dyn. 1–23 (2016). doi:10.1007/s00382-016-3049-z
23. Water, M.: Port Phillip and Westernport Regional River Health Strategy. Yarra
catchment, Richmond (2013)
24. Moghim, S.: Bias Correction of Global Circulation Model Outputs Using Artificial
Neural Networks (Doctoral dissertation, Georgia Institute of Technology) (2015)
25. Sivapragasam, C., Vanitha, S., Muttil, N., Suganya, K., Suji, S., Selvi, M.T., Sudha,
S.J.: Monthly flow forecast for Mississippi River basin using artificial neural net-
works. Neural Comput. Appl. 24(7–8), 1785–1793 (2014)
26. Tschöke, G.V., Kruk, N.S., de Queiroz, P.I.B., Chou, S.C., de Sousa Junior, W.C.:
Comparison of two bias correction methods for precipitation simulated with a
regional climate model. Theore. Appl. Climatol. 127, 1–12 (2015)
27. Teutschbein, C., Seibert, J.: Bias correction of regional climate model simulations
for hydrological climate-change impact studies: review and evaluation of different
methods. J. Hydrol. 456, 12–29 (2012)
28. Um, M.J., Kim, H., Heo, J.H.: Hybrid approach in statistical bias correction of
projected precipitation for the frequency analysis of extreme events. Adv. Water
Resour. 94, 278–290 (2016)

Vous aimerez peut-être aussi