asas t PowerPoint

1
EVALUATION OF PRECIPITATION SPATIAL
INTERPOLATION METHODS IN PINIOS RIVER BASIN,
GREECE ATHANASIOS LOUKAS Department of Civil
Engineering, University of Thessaly, Pedion
Areos, 38334 Volos, Greece E-mail
aloukas_at_uth.gr LAMPROS VASILIADES Department of
Management of Environment Natural Resources,
University of Thessaly NICOLAS R.
DALEZIOS Department of Agriculture, University of
Thessaly
2

• AIM OF THE STUDY
• To evaluate several spatial interpolation methods
  for mean monthly and annual precipitation mapping
  in Pinios River basin, Greece
• OBJECTIVES
•  
• to evaluate a suitable model for mean monthly and
  annual precipitation mapping using
  cross-validation and independent validation
• to estimate gridded monthly and annual
  precipitation data with resolution of 1 Km for
  Pinios River basin
• to evaluate the ability of using neural networks
  for precipitation mapping

3

• STUDY AREA
• Pinios River basin
• Area 9448 Km2
• Elevation ranges from 0 m to 2761 m
• Mean elevation 450 m
• Thessaly plain (agricultural region of an area of
  4000 Km2) is traversed by Pinios River
• Mean annual precipitation ranges from 400 mm to
  1850 mm
• The mountain areas receive significant snowpacks
  during the winter months

4
PINIOS RIVER BASIN
5
DATABASE

• 66 precipitation stations were used for the
  period Oct. 1960 to Sept. 2002
• At least 20 years consecutive monthly
  precipitation data
• Regressions from the nearest stations to fill the
  gaps in the precipitation for the period of
  analysis
• Supervised split sample test in order to ensure
  that the study area is covered by precipitation
  stations
• 73 were held for development and 27 of the
  available precipitation stations withheld for
  validation of the methodologies
• 48 precipitation stations for development and
  cross-validation
• Elevation ranges from 31 m to 1183 m (mean
  elevation 559 m)
• Mean annual precipitation from 409 mm to 1850 mm
  (mean 878 mm)
• 18 precipitation stations for independent
  validation
• Elevation ranges from 95 m to 1400 m (mean
  elevation 505 m)
• Mean annual precipitation from 470 mm to 1450 mm
  (mean 751 mm)

6
ELEVATIONAL DISTRIBUTION OF DEVELOPMENT AND
VALIDATION STATIONS
7
LOCATION OF THE PRECIPITATION STATIONS
8

• INTERPOLATION METHODS FOR PRECIPITATION
• Global methods
• Global Regression using spatial coordinates and
  elevation
• Neural networks using spatial coordinates and
  elevation
• Neural networks using spatial coordinates,
  elevation and monthly weighting
• Local Deterministic methods
• Inverse Distance weighting
• Thin Plate Splines
• Geostatistical methods
• Ordinary Kriging
• Universal Kriging
• Combination methods
• Residual Ordinary Kriging
• Residual Inverse Distance Weighting

9

• GLOBAL METHODS
• Global Regression between spatial coordinates and
  elevation (MLR)
• Stepwise Multiple Linear Regression
• X, Y, Z, X2, Y2, Z2, XY, XZ, YZ
• 13 regression models
• Two Artificial Neural Network (3-Layer) models
  were built
• Neural networks for each monthly and annual
  precipitation values (ANN)
• spatial coordinates and elevation (X, Y, Z)
  (input layer)
• Development of 13 models for monthly and annual
  precipitation (output layer)
• Generalized neural network model for all monthly
  precipitation values (GANN)
• Input layer spatial coordinates and elevation
  (X, Y, Z) and monthly weighting (12 dummy
  variables 0 or 1)
• Development of 1 model for monthly precipitation
  (output layer)
• Annual precipitation sum of estimated monthly
  precipitation

10

• ANN ARCHITECTURE
• Feed Forward ANN
• Trained by back-propagation algorithm
• Geometry optimized by trial and error
• Criteria for ensuring good generalization and
  avoiding over-fitting Number of weights lt Number
  of training samples
• Operated in batch mode
• INITIAL WEIGHTS -a, a, ,
  fi number of input nodes
• TRANSFER FUNCTION Sigmoid transfer function for
  adjustment of weights
• ERROR FUNCTION Mean Square Error
• LEARNING RATE Fixed at 0.005
• MOMENTUM TRANSFER 0.8 (lt1 for convergence)

11
GANN ARCHITECTURE
12
DEVELOPMENT RESULTS - GLOBAL METHODS
GANN 1 model
ANN 13 models
MLR 13 models
13

• LOCAL DETERMINISTIC METHODS
• Inverse Distance weighting (IDW)
• Use of 10 nearby stations
• Include at least 5 precipitations stations
• 13 models (12 monthly and one for annual
  precipitation)
• Thin Plate Splines (Splines)
• Use of 10 nearby stations
• Include at least 5 precipitations stations
• 13 models (12 monthly and one for annual
  precipitation)
• Interpolation Assessment
• Cross Validation (jack-knifing or method of
  fictitious point)

14

• GEOSTATISTICAL METHODS
• Ordinary Kriging (OK)
• Fitted theoretical semivariogram models
• Seven Gaussian (Oct, Nov, Dec, Jan, Feb, Mar,
  Year)
• Six Spherical (May, Apr, May, Jun, Jul, Aug)
• Universal Kriging (UK)
• Detrended for spatial coordinates (2nd order
  trend)
• Ordinary kriging for residuals
• Fitted theoretical semivariogram models
• Four Gaussian (Mar, Apr, Jun, Sep)
• Nine Spherical
• Interpolation Assessment
• Cross Validation (jack-knifing or method of
  fictitious point)

15

• COMBINATIONAL METHODS
• Residual Ordinary Kriging (Res OK)
• Detrended for spatial coordinates and elevation
  (MLR with X, Y, Z)
• Ordinary kriging for residuals
• Fitted theoretical semivariogram models
• Nine Gaussian Models
• Four Spherical Models (Mar, Jul, Aug, Sep)
• Residual Inverse Distance Weighting (RIDW)
• Detrended for spatial coordinates and elevation
  (MLR with X, Y, Z)
• Inverse distance weighting for residuals
• Interpolation Assessment
• Cross Validation (jack-knifing or method of
  fictitious point)

16
SEMIVARIOGRAM DEVELOPMENT RESULTS GEOSTATISTICAL
AND COMBINATIONAL METHODS IGF Values (Pannatier,
1996)
17

• INTERPOLATION ASSESSMENT
• Cross-validation statistics of the development
  stations (N48)
• Coefficient of determination, R2
• Deviation of regression line (yax) from 11 Line
• Root Mean Square Error, RMSE
• Mean Absolute Error, MAE
• Independent validation statistics of the
  validation stations (N18)
• Coefficient of determination, R2
• Deviation of regression line (yax) from 11 Line
• Root Mean Square Error, RMSE
• Mean Absolute Error, MAE

18
INTERPOLATION RESULTS CROSS-VALIDATION RMSE
19
INTERPOLATION RESULTS CROSS-VALIDATION MAE
20
INTERPOLATION RESULTS INDEPENDENT
VALIDATION RMSE
21
INTERPOLATION RESULTS INDEPENDENT
VALIDATION MAE
22

• CONCLUDING REMARKS
• Various spatial interpolation methods were
  assessed for mapping of mean monthly and annual
  precipitation in Pinios River basin
• Overall, RIDW, Res OK, UK, and GANN models gave
  the best results as indicated by independent
  validation statistics
• The GANN method uses only one generalized model
  for all months
• Thirteen (13) models should be developed for mean
  monthly and annual precipitation for the other
  methods
• The GANN method could be efficiently used for
  spatiotemporal modeling of monthly precipitation
  for hydrological modeling of Pinios River Basin
  (future work)