You are on page 1of 8

Journal of Water Resources and Ocean Science

2014; 3(2): 22-29
Published online May 30, 2014 (http://www.sciencepublishinggroup.com/j/wros)
doi: 10.11648/j.wros.20140302.12

Monthly predicted flow values of the Sanaga River in
Cameroon using neural networks applied to GLDAS,
MERRA and GPCP data
SIDDI Tengeleng1, 2, 4, NZEUKOU Armand1, *, KAPTUE Armel5,
TCHAKOUTIO SANDJON Alain1, 3, SIMO Théophile6, Djiongo Cedrigue6
1
Laboratoire d’Ingénierie des Systèmes Industriels et de l'Environnement (LISIE), Institut Universitaire de Technologie Fotso Victor,
Université de Dschang, P.O Box 134 Bandjoun, Cameroun
2
Higher Institute of the Sahel, University of Maroua, P.O. Box 46 Maroua, Cameroon
3
Laboratory for Environmental Modeling and Atmospheric Physics (LAMEPA) Department of Physics University of Yaounde 1 P.O Box
812 Yaounde, Cameroon
4
Laboratory of Mechanics and Modeling of Physical Systems (L2MSP) - University of Dschang, Cameroon
5
Geographic Information Science Center of Excellence, South Dakota State University, Brookling, SD 57007, USA
6
Laboratory of Automatic and Applied Informatics (LAIA), IUT-FV, University of Dschang

Email address:
siddit2000@yahoo.fr (SIDDI T.), armand.nzeukou@gmail.com (NZEUKOU A.), armel.kaptue@sdstae.edu (KAPTUE A.),
stchakoutio@yahoo.com (TCHAKOUTIO S. A.), sitheo1@yahoo.fr (SIMO T.) cedrigueboris.djiongokenfack@gmail.com (Djiongo C.)

To cite this article:
SIDDI Tengeleng, NZEUKOU Armand, KAPTUE Armel, TCHAKOUTIO SANDJON Alain, SIMO Théophile, Djiongo Cedrigue.
Monthly Predicted Flow Values of the Sanaga River in Cameroon Using Neural Networks Applied to GLDAS, MERRA and GPCP Data.
Journal of Water Resources and Ocean Science. Vol. 3, No. 2, 2014, pp. 22-29. doi: 10.11648/j.wros.20140302.12

Abstract: The aim of our study is to predict the discharge rate of the river Sanaga using neural network techniques. Our
investigations have taken place in the Sanaga watershed area in Cameroon. The measurement station is situated in the locality
of Edea-Song-Mbengue (04°04’15”N, 10°27’50”E) where we have obtained monthly values of the river Sanaga discharge
rates that have been measured in situ from January 1989 to December 2004. We have trained neural networks (NN), each with
data of parameters such as the surface albedo, the total cloud fraction, the evaporation, the outgoing longwave radiation, the
air temperature, the specific humidity, the surface runoff and the precipitation height. The precipitation values have been
obtained from GPCP (Global Precipitation Climatology Project) and those of the other parameters from the data assimilation
systems GLDAS (Global Land Data Assimilation System) and MERRA (Modern Era-Retrospective analysis for Research
and Application). As desired outputs of the NN during the learning process, we have used the measured river runoff values.
After introducing temporal delays of 01 and 02 months in the learning-process, we could observe the presence of the memory
effect of the parameters used on the temporal evolution of the river discharge rate. After analysis of the performance's criteria
of the NN with the help of the calculated Root Means Square Errors (RMSE) and determination coefficients between
predicted values and in situ observed ones, we have perceived that the NN which takes into account the two-month delay can
predict the river discharge rate with a strong correlation.
Keywords: River Runoff, GLDAS, GPCP, MERRA, Neural Network, Sanaga Watershed area

nowadays evaluated by modern measurement instruments
1. Introduction [1-3]. However, the acquirement, the maintenance and the
In sub-Saharan Africa in general and in Cameroon, in operability of these logistics lead to significant costs that
particular, electrical energy is mostly produced by constitute a handicap in evaluating this parameter in
hydroelectric dams. The mastering of the temporal sub-Saharan Africa. Collecting direct measurement of
evolution of the discharge rate of the rivers concerned stream flow (discharge) on a continuous basis is
allows better adapting the generators used. The river challenging, especially during large flood events [4]. A
discharge rate is a hydrological parameter that can be common practice is to convert records of water-stages into

be influenced by some atmospheric parameters and soil properties like the surface albedo. In a given neural network. To the 2. The aim of this work is to use the neural network techniques to predict the discharge rate of the river Sanaga in Cameroon. The neural network method has been determine the relationship stage-discharge and to control its developed and implemented by many authors and can be variation [6]. it is the physical methods that need fastidious computation steps important to regularly measure the river runoff in order to for each inversion [16]. Scientists like Deming et al. 3(2): 22-29 23 discharges by using a pre-established stage-discharge behind the problem that has to be solved nor a step by step relationship. The nodes are some remote sensed information [8-9]. each possessing a connection weight wi. These values could gathered in input. In the following sections.⋯. b) an example of a neural network with one input layer (02 inputs). Once factors are able to modify temporarily or definitely the trained. Such relationships are often referred to as a conception of a detailed model describing this physical rating curve. Whereby In opposition to traditional methods. the surface runoff and the precipitation height. using radiometric data. some statistically permit a large coverage of possible cases. managing the problems relied to water such as floods and The structure of a neural network can be defined as an droughts [7]. Lek et al. Thus. the total cloud fraction. the stage-discharge relationship phenomenon. If X is the input vector.w. the evaporation.and output  n   n  ( ) (1) data and determines the weights that have to be attributed to y (X ) = g  ∑ w i ⋅ xi + w 0  = g  ∑ w i ⋅ xi  = g WT ⋅ X ɶ each connection between the neurons in order to  i =1   i =0  approximate with accuracy this relationship [15]. the neural network does need neither a detailed comprehension of the physics x0 =1. the neuron receives and adds n inputs xj.x . Tesch and Randeu [13] and Laurence et al. hidden. [12] have used the neural model to establish the relationship river runoff-rain. oriented graph whose nodes represent the neurons and the The runoff values of a river can be estimated after using bows the different connexions between them.wn) and X (2) 0 1⋯ 1 n . the Figure 1. the air temperature. It is also important to emphasize that the considered nowadays as a powerful tool for estimation in forecast of river discharge in a given region allows better several scientific research domains [17-18]. two hidden layers (03 neurons in each) and one methodology required by the developed neural networks output layer (01 neuron) and will then end by presenting the obtained results. [14] have used them to predict the river discharge rate for short-term. Unfortunately. Brief Description of Functioning result a bias w0 is added. the data used. GLDAS and GPCP. W=( w. and output layers (see figure 1). we will briefly give the functioning principle of the neural networks and present thereafter the study zone which is the watershed of Sanaga basin. the specific humidity. the neurons that are connected to others by weights (synapses). precisely at the measurement point Song-Mbengue. the algorithms possess a higher execution speed as water flow and consequently the rating-curve [5]. Only a general comprehension of the is not always a simple unique relationship. The final obtained result is applied Principle of Neural Networks to a transfer function (or activation function) g in order to obtain the output y. a) a neuron model. [11] have developed a neural network named « Gated Expert » (GE) in order to retrieve liquid water content of the atmosphere over the ocean. The estimation of atmospherical and hydrological parameters with the help of neural networks has been in the past and is still now an interesting subject of scientific research. through assimilated data issued from data bases MERRA. it g : IR n → IR learns the statistical relationship between input. Moreau et al. T ɶ = (1. Furthermore. During the learning phase. the output can be The neural network (NN) is formed by simple calculated as follows: computation elements. the outgoing long wave radiation. Journal of Water Resources and Ocean Science 2014. [10] have used them to validate atmospherical profiles of temperature.x )T . The rating-curve observation and measure technique is required in order to has to be actualized more times in accordance with the determine the most appropriate input parameters that number of the possible runoff values.

c) Illustration of the watershed area of Sanaga (measure point in Song-Mbengue : 04°04’15”N.5 S td param eter v alu es 1 HD 0.0692 0.60E-05 385.5 Surf -2 runoff Figure 2.51°N. GPCP (Global Precipitation Climatology 17th east longitude (see figure 2a).0572 4. but the From MERRA. During System) [19-22].dec 2004 2 T b) 1. we have monthly values of the discharge rate of Sanaga (m3/s) obtained in situ in the measurement station 3. b) -2.33°N.10808 0.5 Cloud 0 -0. The general characteristics of the monthly parameter means are represented in the figure 3. there is water in the Sanaga.: Monthly Predicted Flow Values of the Sanaga River in Cameroon Using Neural Networks Applied to GLDAS. 10.14E-07 σ (std dev) 1. Table 1. standard deviation and variation coefficients. (6.5 Month localization of watershed area of Sanaga in Cameroun ((3. We have also used the data assimilation systems Cameroon is situated in central Africa.5 Jan Feb March April May June Juil Aug Sept Oct Nov Dec Evap -1 -1.035 T HD H Surf runoff Statistical parameters [K] [kg/kg] [mm] [kg/m2/s] µ (mean) 289. total cloud fraction. humidity. Figure 3.1979 5. the whole year. b) temperature. we have focused on the watershed area of radiation.dec 2004 1. coordinates (3. 2 Standard values in Edea-Song-Mbengue: jan 89 .826 4. we have obtained the values of the surface discharge rate of this river varies from one year to another.00402 0. They concern the standardized monthly mean values from January 1989 to December 2004.5 a) Albedo 1 Std p ara m e te r v alu es 0. cloud fraction. Evolution of standardized monthly parameter values in the watershed area of the river Sanaga: a) albedo. The country is Project) and GLDAS (Global Land Data Assimilation characterized by a dense hydrographical network. Used Data and OLR (Outgoing Long Radiation). Sanaga Watershed area.57 CV (var coef) 0.00E-06 13. accumulated precipitation and surface runoff. From GPCP are issued the monthly accumulated rain. runoff.6255 0.42°E)). specific humidity and surface Sanaga that is approximately delimited by the geographical runoff.5 0 -0.10°E) and (6. we have obtained the surface figure 2b). standard deviation and variation coefficient concerning the period from January 1989 to December 2004 in the region of the Sanaga watershed area (Cameroon) Albedo CLDTOT EVAP OLR Statistical parameters [fraction] [fraction] [kg/m2/s] [W/m2] µ (mean) 0.1.16E-07 CV (var coef) 0.162 0.42°E) (see and from the GLDAS system.1. 14. between January 1989 and December 2004.10°E). albedo. 14.123 0. outgoing longwave As study zone. Table 1 presents the statistical values of the mean.51°N. Study Zone and Used Data At disposal. a) localization of Cameroon in Africa. 10°27’50”E).01519 140.001641 88. Study Zone of Song-Mbengue.993 6. evaporation. MERRA and GPCP Data 3. and between the 8th and the and Application).840 .02 σ (std dev) 0.5 Sdandards values in Edea-Song-Mbengue.087 0.5 -2 OLR -2.5 Jan Feb March April May June Juil Aug Sept Oct Nov Dec H -1 -1.047 0. jan 89 .30 0. between the 1st and MERRA (Modern Era-Retrospective analysis for Research the 13th degree north latitude. air temperature. evaporation 3. 10.5 Month 2.33°N. Statistical parameter values: mean.24 SIDDI Tengeleng et al.0058 0.

the evaporation (EVAP) . up to now. Xmin and Xmax represent respectively the real process is stopped. the learning Whereby X. Methodology iteration number. and the NN can be considered as value of the parameter. be condensed in 05 main points: Sometimes. learning process. it is necessary to work with (sample).and output values of existing hidden layers are purpose.and output values of the last layer (output layer) X − X min are computed. during the matrix. and the outputs presented to him. Many stop-criteria exist. and a new iteration begins. If the evaluated error is important. its minimal value and its maximal trained and used as a prediction tool for other outputs value. the next the cloud coverage (%) (CLDTOT). If the calculated error is not significant. but the main important ones are the 4. it is necessary. X norm = 2 ⋅ −1 (3) 4. Learning process of a neural network For a convenient prediction of the outputs corresponding input vector (sample) is read from the available input to given inputs of a trained NN. In the figure 4. we have presented a employed the neural network toolbox available in MATLAB diagram describing the learning process of a NN which can [23-24]. normalized input values situated between -1 and +1. ends if the NN has sufficiently learnt the relationship Figure 4. parameters (equation 3): 3. From the input matrix. The error between desired available value and NN X max − X min output is evaluated. Journal of Water Resources and Ocean Science 2014. Input. that all possibilities of inputs / outputs For the learning of the different parameters used. in accordance with different stages of the 1. we have used the normalized NN input values of calculated. Xpi is an input vector xpi read parameter values used. 5. and the error calculated between the Training of a neural network is an iterative process that desired values and those evaluated by the NN. unknown As inputs parameters of the NN we have: the albedo (%). we have combinations exist. corresponding to inputs that are. by the NN. 3(2): 22-29 25 existing between the inputs. Input. For this 2.

011 the second the number of neurons in the first layer. to December 1997.95 fourth the number of outputs of the NN. the 1000 second the number of neurons in the third layer and the 0 r = 0. 5. we have compared in figure 5 the predicted and the measured river ∑ (x ) N n − x is 2 discharge rate values. RM SE = i =1 N −1 7000 Training phase 02 months delay 8-10-10-1 RN prediction Measure ∑ (x )⋅(x ) N 6000 −x −x (a) n n s s 3 /s) 5000 i i Sanaga runoff in Song-Mengue (m^ r= i =1 (5) 4000 ∑ (x ) ∑ (x ) N 2 N 2 n −x n ⋅ s −x s 3000 i i 2000 i =1 i =1 1000 n s Whereby xi and xi are respectively the NN simulated 0 values.1.96*x + 0. With the 02 months’ time delay. the in situ measured on the evolution of the temporal river discharge rate. the air With the first NN. 7000 02 months delay Training phase 6000 5. from gives the different periods corresponding to the training and February 1989 and from March 1989. 8-10-20-1. MERRA and GPCP Data (kg/m2/s). We have observed -1000 0 1000 2000 3000 4000 5000 6000 7000 that the structure 8-10-10-1 is better trained as the other Measure (m3/s) structures.2. the discharge rate in Song-Mbengue. The second RUNOFF) (kg/m2/s). Training Phase of the Neural Networks After execution of the input values through the trained We have trained for each temporal delay three NN with 02 NN we can observe that the 02 months delay leads equally to hidden layers (8-10-10-1.: Monthly Predicted Flow Values of the Sanaga River in Cameroon Using Neural Networks Applied to GLDAS. Results and Discussions 5000 (b) N N p red ic ted (m3 /s ) 4000 We have trained for each temporal delay three NN with 02 hidden layers (8-10-10-1. The temporal delays that we have introduced allow 1996.95 orders. Table 2 river discharge rates respectively from January 1989. Table 2. 8-20-10-1). All these NN input values 2004) and 02 months (as from March 1997 to September correspond to the period from January 1989 to December 2004). neural networks during the prediction of the river Sanaga the second the number of neurons in the first layer. 10 neurons in the discharge rate values. We have observed Square Error (RMSE) (equation 4) and determination that the structure 8-10-10-1 is better trained than other coefficient (r) (equation 5): structures. In this better predictions. 8-20-10-1). Prediction Phase of the Neural Networks 5. during the learning process. without any accumulated rain (H) (mm) and the surface runoff (SURF time delay accordingly to the training period. Training of the neural network 8-10-10-1 with a time delay of 02 compared in figure 5 the predicted and the measured river months (8 input parameters. 8-10-20-1. In this 3000 designation the first number represents the number of inputs. the outgoing long radiation (OLR) (W/m2). Training and prediction periods for the three NN NN Training period Prediction period 1 January 19889 – December 1996 January 1997 – November 2004 2 January 19889 – December 1996 February 1997 – October 2004 3 January 19889 – December 1996 March 1997 – September 2004 For better understanding the efficiency of the different designation the first number represents the number of inputs. the in situ measured ones and N the total number of -1000 0 10 20 30 40 50 Training Period: jan 89 . the rates as from January1997 to November 2004. we have predicted the river discharge temperature T (°K). As outputs of the three NN. With the 02 months’ time delay. Thus. the specific humidity (HD) (kg/kg). 10 neurons in the 1st hidden. to the prediction phases. We have found a strong correlation of 2nd hidden layer (see the legend) 0.95 orders. we obtain for the period from March 1997 to September 2004 the predicted river discharge . We have found a strong correlation of i (4) 0.26 SIDDI Tengeleng et al. 2000 y = 0. We have obtained the values from the and the third NN predict these values respectively with a NASA website and through the data assimilation systems temporal delay of 01 (as from February 1997 to October MERRA. we have Figure 5. we have evaluated the second the number of neurons in the third layer and the performance criteria through computation of the Root Mean fourth the number of outputs of the NN.dec 96 60 70 80 90 100 samples. we have used as desired perceiving the memory effect of the used input parameters values. GPCP and GLDAS.

88*x + 0.0 0.sept 04 Sanaga runoff in song-Mbengue (march 97-sept 2004) 7000 Predition phase 02 months delay NN prediction vs in situ measure 6000 linear regression (b) 5000 N N p re d ic ttio n (m3 /s ) 4000 3000 2000 y = 0.83 2 430. Root Mean Square Errors (RMSE) and Determination Coefficients (r) between the discharge rate values. calculated by the NN and measured in situ: training phase from January 1989 to December 1996. prediction phase from January 1997 to December 2004. b) Linear regression between measured and predicted values. a) Prediction with the 8-10-10-1 NN (squares) and measures (points) of the Sanaga discharge rate at Song-Mbengue in Cameroon during the period from March 97 to September 2004.50 0. order of 0.0 0.92 17 882. 7000 Predition phase NN predicted 02 months delay in situ measured 6000 (a) S an a g a ru no ff in S on g -M b e ng u e (m3 /s ) 5000 4000 3000 2000 1000 0 0 10 20 30 40 50 60 70 80 90 100 Measure period : March 97 . 01 months and 02 months delay) Training phase Prediction phase Delay time (months) RMSE Iteration RMSE r r (m3/s) number needed (m3/s) 0 439. 3(2): 22-29 27 values of Sanaga at the measurement point Song-Mbengue We obtain equally the existence of a strong correlation in that we compare in figure 6 to the values observed in situ. The table 3 presents the different values of the performance criteria RMSE and r obtained Table 3.95 18 735.9 0.41 0.80 1 434. Journal of Water Resources and Ocean Science 2014.89 0 0 1000 2000 3000 4000 5000 6000 In situ measure (m3/s) Figure 6.00 0. use of the configuration 8-10-10-1 for all the three temporal delays (0 month.037 1000 r = 0.91 12 1056.89 .89 between predicted and measured values.

2002. Barbetta. measured ones). 319-331. “Atmospheric Liquid Water Retrieval Using a Research and Application Gated Experts Neural Network”. P. R. 86. Eau. “Estimating discharge rate in rivers CLDTOT: Cloud coverage using remotely sensed hydraulic information”. and Randeu. Revue des sciences de RMSE: Root Mean Square Error l’eau. Volume of Surface Runoff from in Situ Measured Soil Sorptivity”. doi:10. Y. B. criteria (RMSE and r between predicted values and in situ No. C. Earth neural network technique has been used after the training of Syst. the use agricultural catchments”. Angelica. MERRA and GPCP Data 6. efficient and rapid mean with low cost that permits to predict [6] Watanabe... Silvia. I. Hazards Earth Syst. Ghachtoul. “Space-Based Measurement of River Runoff Transactions”. M. L. Suzuki. Conclusion In our study. Smith.. EOS: Earth Observing System [9] Tarpanelli. 62-79. M. Dimopoulos. ASCE/May 2006. 629-635. Furthermore.. 2005.. C. Thiria. Journal of CV: variation coefficient Hydrology.3390/rs50941 45 GLDAS: Global Land Data Assimilation System [10] Deming J.. Meselhe.. Neural Networks. Chaohua D. 2003. 60641L. discharge from earth observation measurements of river While developing the NN. N. C. Sci. we have investigated the prediction of the References discharge rate of the river Sanaga in Cameroon through [1] Chen. American Geophysical. world during the 19th – 20th centuries”.: Monthly Predicted Flow Values of the Sanaga River in Cameroon Using Neural Networks Applied to GLDAS. and the configuration (8-10-10-1) has been identified 15. Earth Syst.649743. Journal of Arid Land Studies. the models with values observed in the measurement station [2] Negrel. Kobayashi. Luca and EVAP: Evaporation Moramarco. Sci. and MERRA: Modern Era-Retrospective analysis for Klapisz.. Y. 6. 2006. after analyzing the performance [3] EOS. MERRA and GPCP appears as an adequate. 309. several architectures have been surface hydraulic variables”. “A neural network model for Surf runoff: Surface runoff short term river flow prediction”. no..1117/12. Information Services Center Remote Sens.. SPIE 6064. Acknowledgments [7] Pekarova. Habib and Ehab A. 1951–1962. “Estimating river Song-Mbengue. 191–209. Journal of Atmospheric and Oceanic Technology. P. Badran.. Shinji.. rev. “Stage-discharge for better prediction of the river discharge rate with strong Low-Gradient Tidal Streams Using Data-driven Models”. Tommaso. 236. Fumio. and HD: Humidity Machine Learning. Mallet. A mountain river . MERRA and GPCP. tested. 2013. 2011. 17. IAHS Publ. E. Tomoki and Takahashi. 95-98. Laurence.. F. P. S. W. Delwyn. independently on the temporal delay introduced. correlation.. Sci. 2012. Oygarden.. Global of NN models based on data from assimilation systems like and Regional Perspectives Centre for Soil and Environmental Research GLDAS. 1996. 2013. Kosuth. “Flood discharge rate measurement of a using assimilated data from GLDAS. “Estimating the in Cameroon with a good spatio-temporal regularity. “Erosion and surface runoff in small values need enormous material and human resources. Nomenclature [8] Bjerkliea David.Nanshih River in Taiwan”. as the best one. 19. Image Processing: Algorithms and Systems. Journal of Hydraulic Engineering. Considering that the measures of river discharge rate [5] Lillian. 22 (1). NASA: National Atmospheric and Space Administration NN: Neural Network [12] Lek. Nat. S. Proc. Miklanek and Jan Pekar. 5. 19.. 274. the discharge rate values of rivers of Sanaga watershed area Hotta. Hydrol. 3.28 SIDDI Tengeleng et al. Brocca. Union. “Discharge Estimation by Using GES DISC: Goddard Earth Sciences Data and Altimetry Data and Simplified Flood Routing Modeling”. T: Air temperature Sci. 10 May 2005. « Rainfall-runoff OLR: Outgoing Long Radiation modelling using artificial network”. we have perceived that the NN which takes into account the temporal delay of 02 months produces [4] Emad. σ: Standard deviation [13] Tesch.. 2006. 1996. Journal of Hydrology. El. Satoru. H. 2049–2058. and Weison L. Mabbou. HDISC: Hydrology Data and Information System Center µ: Mean value [11] Moreau. 4145-4162. and Bercher. “Neural networks GOES: Geostationary Operational Environmental approach to high vertical resolution atmospheric temperature Satellite profile retrieval from space borne high spectral resolution GPCP: Global Precipitation Climatology Project infrared sounder measurements”. Lawrence Dingman. Yukimitsu. Vol. . Hydrol. “Spatial and We thank NASA through its service data (Goddard Earth temporal runoff oscillation analysis of the main rivers of the Sciences Data and Information Services Center (GES DISC)).. Eos. Pavol. doi:10. 457-467. C.

(9). doi:10. and hydraulic Mitchell. 0975-8887.. Proceedings of the IEEE. M. Rodell. Hiroko. Gupta... L.M. J. H. Karmakar.. [19] Huffman. Jambor. Hydrology Data Applications. 36. data”. Fall networks: Perceptron. B. M. 2009. “Version 1. 1996..May 2. H. K. Smith and Tamlin. Journal of [24] Demuth. P. Atmospheric Processes Laboratory. and Toll. C. “Application of Artificial Neural [21] Hongliang. J. and Information Services Center (HDISC)”. Beate.. 51. K.J. 3(2): 22-29 29 [14] Laurence. “Neural Networks for the retrieval of water vapour and liquid water from radiometric [22] Rui. Bulletin of American Meteorological Society. Radio Science. and Manoj Kumar Kowar. P. “Estimation [20] Rodell. Boston”: Pws publication. “Global Land Data Assimilation Literature Review”. “Neural Network [18] Kou-Lin Shu. M. D. Siberia”. U. S. Gottschalck. Networks in Weather Forecasting: A Comprehensive W. International Journal of Computer System (GLDAS) products from NASA. toolbox user’s guide”. 2013. “Neural Network applied meteorology. K. G. of river discharge rate. Ressources Research. “NASA Giovanni Portals for NLDAS/GLDAS Online Visualization. M. 78. propagation speed.1029/2007 Lohmann. Pavelsky. D. 1990. D. and Giovanni S. Madaline. M. doi: 10.. 381-394. 33... J.. (18). M. M. 2012.. 1176-1189.. Analysis and [17] Widrow. April 28 .K. and Rodell.. Walker. M. Teng. .. “30 years of adaptive neural Intercomparison”. D.. F. M. American Geophysical Union.A.1175/BAMS-85-3-381. J. Cosgrove.R... V. Assimilation System”. Xiaogang Gao. 2008.. Meeting. [16] Del Frate. L... J. G.K. 2011. Guhathakurta. Inc. Beaudoing. and Hogan. geometry from space: Lena River. F. 1373-1386. [23] Hagan. Mesoscale”.. 1998. Entin. [15] Shrivastava.. and Beate.E. Lehr..).. Bolvin.2 GPCP One-Degree Daily Precipitation Data Set Documentation. The Mathworks. B.E. 1996. and Backpropagation”. Bosilovich. M. H. (5). Natrick... 2006. H. Water Radakovich. L. and Vollmer. Teng W. Demuth.. Vollmer B. Hrubiak. Vol.. Journal of Water Resources and Ocean Science 2014. M. Arsenault. Sorooshi Sorooshia and Hoshin design. Mocko. “Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks”.. B.P. USA. T. C. 2004.. Portland Oregon. ASPRS Annual Conference.K.. Meng. P. “The Global Land Data WR006133. NASA Goddard Space Flight Center. 44. Houser.