See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/253576027

A regression-kriging model for estimation of rainfall in the Laohahe basin

Article in Proceedings of SPIE - The International Society for Optical Engineering · October 2009
DOI: 10.1117/12.838036

CITATIONS READS
0 569

3 authors, including:

Hong Wang
Hohai University
26 PUBLICATIONS 1,068 CITATIONS

SEE PROFILE

All content following this page was uploaded by Hong Wang on 03 September 2014.

The user has requested enhancement of the downloaded file.

 A regression-kriging model for estimation of rainfall in the Laohahe
basin
Hong Wang* a, Li L. Ren a, Gao H. Liu b
a
State Key Laboratory of Hydrology, Water Resources and Hydraulic Engineering, the College of
Hydrology and Water Resources, Hohai University, Nanjing, China 210098; bInstitute of Geographic
Sciences and Natural Resources Research, CAS, Beijing, China 100101

ABSTRACT

This paper presents a multivariate geostatistical algorithm called regression-kriging (RK) for predicting the spatial
distribution of rainfall by incorporating five topographic/geographic factors of latitude, longitude, altitude, slope and
aspect. The technique is illustrated using rainfall data collected at 52 rain gauges from the Laohahe basis in northeast
China during 1986-2005 . Rainfall data from 44 stations were selected for modeling and the remaining 8 stations were
used for model validation. To eliminate multicollinearity, the five explanatory factors were first transformed using factor
analysis with three Principal Components (PCs) extracted. The rainfall data were then fitted using step-wise regression
and residuals interpolated using SK. The regression coefficients were estimated by generalized least squares (GLS),
which takes the spatial heteroskedasticity between rainfall and PCs into account. Finally, the rainfall prediction based on
RK was compared with that predicted from ordinary kriging (OK) and ordinary least squares (OLS) multiple regression
(MR). For correlated topographic factors are taken into account, RK improves the efficiency of predictions. RK
achieved a lower relative root mean square error (RMSE) (44.67%) than MR (49.23%) and OK (73.60%) and a lower
bias than MR and OK (23.82 versus 30.89 and 32.15 mm) for annual rainfall. It is much more effective for the wet
season than for the dry season. RK is suitable for estimation of rainfall in areas where there are no stations nearby and
where topography has a major influence on rainfall.
Keywords: regression-kriging, factor analysis, GLS, rainfall

1. INTRODUCTION
Effective watershed management strategies depend on accurate model results. Rainfall is the most important input for
watershed modeling, including hydrology modeling. Rainfall characteristics are usually spatially varying, even in a small
watershed; so accurate description of the spatial variation of rainfall is quite important for predicting water movement in
a watershed. The accurate estimation of the spatial distribution of rainfall requires a very dense network of instruments,
which entails large installation and operational costs. Therefore, it is essential to estimate point rainfall at unrecorded
locations from values at surrounding sites.
A number of methods have been proposed for the interpolation of rainfall1, including global interpolation methods (trend
surface and multiple regression), local interpolation methods (Thiessen polygons, inverse distance weighting, kriging and
splines), and mixed methods (combined global and local methods). Rainfall generally increases with elevation2-3, and so
many authors have incorporated elevation into its prediction using geostatistical approaches1, 4-5. Others have developed
relationships between rainfall and various topographic variables such as altitude, latitude, continentality, slope,
orientation or exposure, using regression approaches6-11. In recent years, there has been an increasing interest in hybrid
interpolation techniques in which two conceptually different approaches are combined to model and map rainfall: (1)
interpolation relying solely on point observations of rainfall; and (2) interpolation based on regression of rainfall on
spatially exhaustive auxiliary information (such as topographic variables). Several studies have been conducted in a
range of locations and environments around the world (for example, in various European countries, the USA, Australia

*
[email protected]; phone 86 25 83787891; fax 86 25 83592288

International Symposium on Spatial Analysis, Spatial-Temporal Data Modeling, and Data Mining, edited by Yaolin Liu,
Xinming Tang, Proc. of SPIE Vol. 7492, 74924I · © 2009 SPIE · CCC code: 0277-786X/09/$18 · doi: 10.1117/12.838036

Proc. of SPIE Vol. 7492 74924I-1

and China). They showed that the hybrid technique can yield better predictions than either of the individual approaches1,
4-5, 12-14
. One of these hybrid interpolation techniques is known as regression-kriging15-17. It first uses regression on
secondary information and then uses simple kriging (SK) with a known mean (0) to interpolate the residuals of the
regression model. It is mathematically equivalent to the interpolation method variously called “Universal Kriging” (UK)
and “Kriging with External Drift” (KED), where secondary predictors are used directly to solve the weighting problem in
kriging. Originally, UK was intended as a generalized case of kriging in which the trend is modeled as a function of
coordinates within the kriging system. If the drift is defined externally as a linear function of some secondary variables
rather than the coordinates, the term KED is preferred. In case of UK or KED, predictions are made as with kriging, but
the covariance matrix of residuals is extended with the secondary predictors18. However, drift and residuals can also be
estimated separately and then added together. This procedure was later named regression-kriging15. Goovaerts19 used the
term kriging with trend model to refer to a family of interpolators while RK is referred to as simple kriging with varying
local means. Whether KED or RK, the variogram parameters must be estimated from regression residuals. This
regression should be generalized least squares (GLS) because of the likely spatial correlation among residuals. RK has
the advantage of explicitly separating trend estimation from residual interpolation, allowing the use of arbitrarily
complex forms of regression instead of the simple linear techniques that can be used with KED17.
In many cases, kriging combined with regression has proven to be superior to plain geostatistical techniques, yielding
more detailed results and higher accuracy of prediction14. Although there are diverse interpolation methods in existence,
no single method is suitable for every circumstance20. For instance, the straightforward linear regression of rainfall
against elevation generally out-performed OK if their correlation coefficient was larger than 0.751. Lloyd13 concluded
that kriging with an external drift (KED) provides the most accurate estimates of rainfall for all months from March to
December whereas for January and February OK provides the most accurate estimates.
The objective of this study is to evaluate the suitability of RK in predicting the spatial distribution of monthly rainfall in
the Laohahe basin of China, in which five topographic/geographic factors (latitude, longitude, altitude, slope and aspect)
were incorporated. Rainfall data from 44 stations were selected for the prediction and the remaining eight stations were
used for validation. The predicted rainfall distribution was compared with that from the OK and MR techniques.

2. CASE STUDY
The Laohahe basin lies in southeastern Inner Mongolia Autonomous Region, China, with an area of approximately
1.86×104 km2. The 52 rain gauges (Fig. 1) with daily readings were randomly divided into interpolation (44) and
validation sets (8) in this study. The monthly and annual rainfall data have been averaged over the period of January
1986-December 2005. Rainfall is strongly seasonal (Fig. 2), and so in the analysis mean monthly rainfall was calculated
for the dry season (June - September), the wet season (the rest of the year), as well as the whole year. The basic rainfall
statistics (mean, standard deviation, minimum, and maximum) are presented in Table 1.
Another source of information is an SRTM DEM of 90m resolution downloaded from the International Agriculture
Research Consortium for Spatial Information server (http://srtm.csi.cgiar.org). Elevation (ELE), slope (SLP) and aspect
(ASP) were derived from the SRTM DEM that had been resampled to a cell size of 200 m in ArcGIS 9.1. Elevation
ranges from 433 to 2047 m. It has an apparent W-E gradient with the highest in northwest part of the study area as shown
in Fig. 3. There is a significant correlation between rainfall and latitude and slope at the 0.01 level in different periods
(Table 1). However, the opposite is true with longitude and aspect. Apart from the dry season, the correlation between
rainfall and elevation is also significant.

3. METHODS
3.1 Transformation of predictors
Factor analysis was used to produce standardized Principal Components (PCs) prior to regression in order to eliminate
multicollinearity. Unlike the original predictors, these uncorrelated and standardized PCs do not share any collinearity
among them . From the 44 rain stations, five predictors, including ELE, SLP, ASP, X and Y, were transformed and the
first three PCs were used for stepwise regression.

Proc. of SPIE Vol. 7492 74924I-2

3.2 Kolmogorov-Smirnov’s test of rainfall data
The Kolmogorov-Smirnov’s test was used to determine whether rainfall has a normal distribution. If not, it should be
logarithmically transformed to e be sure to model non-linear relationship between rainfall and predictors and to force
predicted values to fall within physical limits8.
3.3 Regression-kriging technique
RK combines two approaches of regression to fit the explanatory variation with SK with the expected value of 0 to fit
residuals, i.e. unexpected variation:
p n
zˆ( s0 ) = mˆ ( s0 ) + eˆ( s0 ) = ∑ βˆk ⋅ qk ( s0 ) + ∑ λi ⋅ e( si ) (1)
k =0 i =0

ˆ ( s0 ) is the fitted drift, eˆ( s0 ) is the interpolated residual, βˆk is the estimated drift model coefficient,
where m λi is the
kriging weight determined by the spatial dependence structure of the residual, and e( si ) is the residual at location si .
The regression coefficient βˆk is estimated from the samples by the GLS method in order to take heteroskedasticity or
spatial correlation between individual observations into account. The GLS coefficient βˆk is computed using the EViews
package21 in following steps:
(1) Determine a linear model of the variable as predicted by the auxiliary map q (here it is two PCs). Then the most
suitable predictors are selected and OLS regression coefficients are obtained;
(2) Derive OLS residuals at interpolation sample locations as

e* ( si ) = z ( si ) − [b0 + b1 ⋅ q( si )] (2)

(3) Test normalcy of data distribution, heteroskedasticity and spatial autocorrelation of the OLS residuals using the
Jarque-Beta, Godlfeld-Quandt and Lagrange Multilier test22 , respectively, in the EViews package;
(4) Calculate weight in the weighted ordinary least square (WLS). In light of heteroskedasticity in the residuals, OLS is
not BLUE (Best Linear Unbiased Estimator). So the original predictors should be transformed in order to satisfy the
condition that the variance is a constant. Then OLS is used for these transformed predictors. This is called GLS.
Generally, GLS is the OLS of transformed variables which satisfy the standard hypothesis of OLS. In order to transform
the original predictors, the relationship between residual variance and predictors should be explored first. The best
regression model was determined using curve estimation in the SPSS package. If the model is linear regression, then the
−2 −1
weight = q ; if the model is quadratic regression, then the weight = q . Both sides of the regression equation are
multiplied by this weight to make the estimated coefficients the GLS coefficients;
(5) Derive GLS residuals and test the normalcy of data distribution, heteroskedasticity and spatial autocorrelation of the
GLS residuals using the same techniques as in step (3);
(6) Interpolate the GLS residuals using SK with a known expected mean of residuals (by definition 0) and transform the
interpolated map into a raster layer of a 200 m resolution in ArcGIS; and
(7) Overlay the GLS surface with the interpolated GLS residuals at each prediction point in ArcGIS. This processing
comprises three steps: (i) the GLS coefficient was used to compute the rainfall in different periods in each 200 m-interval
prediction point; (ii) interpolated rainfall maps were produced using the inverse distance weighting (IDW) method and
resampled to 200 m resolution; and (iii) the GLS surface was added to the interpolated GLS residuals raster layer.
3.4 Evaluation
The performance of RK is assessed and compared with OK and OLS MR using the validation dataset of eight stations.
The true prediction accuracy is evaluated by comparing the estimated values ( zˆ( s j ) ) with the observed ones at
*
validation point ( z ( s j ) ) in order to assess systematic error, calculated as mean absolute error ( MAE ):

Proc. of SPIE Vol. 7492 74924I-3

 1 l
MAE = ⋅ ∑ zˆ( s j ) − z * ( s j ) (3)
l j =1
and the accuracy of prediction, expressed as root mean square error (RMSE):

1 l
RMSE = ⋅ ∑ [ zˆ ( s j ) − z * ( s j )]2 (4)
l j =1
where l is the number of validation points. In order to compare accuracy of prediction between variables of different
type, the RMSE can be normalized by the total standard deviation ( s z )of observed samples:

RMSE
RMSEr = (5)
sz
As a rule of thumb, a RMSEr value close to 40% is considered a fairly satisfactory accuracy of prediction. By
comparison, a value >70% means that the model accounts for less than 50% of variability at the validation points and the
prediction is deemed unsatisfactory.

4. RESULTS AND DISCUSSION

4.1 RK modeling
Table 1 shows that the rainfall data in different periods have a normal distribution as confirmed by the Kolmogorov-
Smirnov test at the 0.05 level. The results of factor analysis indicate that the first three PCs account for more than 90% of
the total variation in the bands (48.3%, 24.2% and 17.6%). PC1 as the main component explains the variations in ELE,
SLP and X. PC2 accounts mainly for the variation in Y, SLP and X, while the third component accounts for ASP. The
last three columns in Table 1 indicate the existence of a significant correlation between rainfall and the first two PCs at
the 0.01 level, but not PC3.
Stepwise regression substantially reduces the number of predictors. In the wet season and over the whole year, the
predictors are reduced to PC1, PC2 and its square, while in the dry season, only PC1 and PC2 are selected (Table 2). The
normality test (Jarque-Beta) for OLS residuals shows that the distribution is far from normal in the wet season, and the
Lagrange-Multilier test indicates the absence of significant autocorrelation in all the periods. The Godlfeld-Quandt test
shows significant heteroskedasticity in the dry season and over the whole year (Table 3). In light of heteroskedasticity in
the residuals, the standard errors of the OLS coefficient estimates are incorrect. In this case EViews was used to obtain
the GLS estimator, which was derived by transforming the predictors to produce a statistical model with homoskedastic
errors. Table 4 shows that the quadratic regression model is the best for the dry season and the annual period. However,
in the dry season the linear model is the best. Thus, in the dry season and over the annual period, both sides of the
estimation equation was multiplied by (PC2)-1; in the wet season, the equation was multiplied by (PC2)-2, in deriving the
GLS coefficient (Table 2). The last three rows in Table 3 show that GLS residuals are normally distributed without
autocorrelation and heteroskedasticity.
4.2 Comparison of predicted results
The results of RK, OK and MR (OLS multiple regression) in different periods over the study area are presented in Fig.4.
The OK maps (Fig. 4a_1, a_2 and a_3) show rather steep transitions with a fairly low level of details, whereas the MR
maps (Fig. 4b_1, b_2 and b_3) reflect change in latitude, elevation and slope. The RK (Fig. 4c_1, c_2 and c_3) and OK
maps highly resemble each other, and yet the RK maps yield more details than the OK maps, in addition to showing hot
spots (obvious in Fig. 4c_2) invisible in the MR map (Fig. 4b_2).

Proc. of SPIE Vol. 7492 74924I-4

4.3 Spatial pattern of rainfall
In the wet season, rainfall clearly decreases from the south to the north. Thus, latitude plays a very important role in this
period. The rainfall distribution for the dry season is highly similar to that of the DEM (Fig. 3). Hence, altitude or slope
plays a leading part. Over the whole year, there is a strong N-S gradient with less rainfall in the northeast and relatively
more rainfall in the south and northwest. Obviously, the spatial distribution of rainfall in the Laohahe basin is influenced
mainly by latitude, while slope and elevation are secondary influencing factors (Table 1).
4.4 Bias and accuracy of prediction
The summary comparison of the prediction methods in different periods is presented in Table 5. Over the whole year,
RK achieves a higher prediction accuracy (44.7%) than OK (73.60%) and MR (62.27%) and a smaller bias (23.82 versus
32.15 and 30.89 mm). This relativity remains unchanged in both the wet and dry seasons. The relatively low bias for RK
in all cases demonstrates that the use of PCs and GLS regression coefficient estimation serves their purpose quite well. In
the dry season, both MR and OK have a relative prediction error of larger than 70% or explain less than 50% variation.
The reason is that in this season the study area is under the influence of completely different synoptic systems, especially
the strong Siberian high pressure system, and intensified westerly winds23. They provide little rainfall in the Laohahe
basin and disturb the general pattern of rainfall distribution, thus affecting the prediction accuracy of the models.

5. CONCLUSIONS
Clearly, RK is a powerful prediction technique that can be used to interpolate sampled environmental variables (such as
elevation, slope, aspect, longitude and latitude) from large point datasets. Factor analysis on the predictors is efficient at
removing multicollineality and reducing asymmetry in their distribution. GLS coefficient estimation ensures normality,
randomness and homoskedasticity of the prediction residuals. RK achieves a higher prediction accuracy than OK and
MR (44.67% versus 73.60% and 62.27%) and produces a smaller bias (23.82 versus 32.15 and 30.89 mm) for annual
rainfall. It is much more effective for the wet season than for the dry season. RK enables the estimation of rainfall in
areas where there are no stations in the vicinity and where topography exerts a major influence on rainfall. Further
research should investigate whether other environmental descriptors, such as the Tropical Rainfall Measuring Mission
(TRMM) 24-25, the Normalized Different Vegetation Index (NDVI) 26 or local prevailing winds (especially their direction
and force) are able to account for a larger proportion of the spatial variability displayed by rainfall.

ACKNOWLEDGEMENTS

This research was supported financially by the Natural Science Foundation of China (Grant No. 40871230) and the
National Key Basic Research Program of China (Grant no. 2006CB400502).

REFERENCES

[1] Goovaerts, P., “Geostatistical approaches for incorporating elevation into the spatial interpolation of rainfall”, J.
Hydro., 228, 113-129 (2000).
[2] Spreen, W. C., “A determination of the effect of topography upon precipitation”, Trans. Am. Geophys. Union, 28,
285-290 (1947).
[3] Smith, R. B., [The influence of mountains on the atmosphere], Adv. Geophys, 21, Academic Press, 87-230 (1979).
[4] Martínez-Cob, A., “Multivariate geostatistical analysis of evapotranspiration and precipitation in mountainous
terrain”, J. Hydro., 174, 19-35 (1996).
[5] Prudhomme, C. and Duncan, W. R., “Mapping extreme rainfall in a mountainous region using geostastistical
techniques: a case study in Scotland”, Int. J. Climatol, 19, 1337-1356 (1999).
[6] Basist, A., Bell, G. D.and Meentemeyer, V., “Statistical relationships betweent opography and precipitation patterns”,
J. Clim., 7, 1305-1315 (1994).

Proc. of SPIE Vol. 7492 74924I-5

[7] Goodale, C. L., Alber, J. D.and Ollinger, S. V., “Mapping monthly precipitation, temperature and solar radiation for
Ireland with polynomial regression and digital elevation model”, Clim. Res., 10, 35-49 (1998).
[8] Ninyerola, M., Pons, X. and Roure, J. M., “A methodological approach of climatological modeling of air temperature
and precipitation through GIS techniques”, Int. J. Climatol., 20, 1823-1841 (2000).
[9] Wotling, G., Bouvier, Ch., Danloux, J. and Fritsch, J. M., “Regionalization of extreme precipitation distribution using
the principal components of the topographical environment”, J. Hydro., 233, 86-101 (2000).
[10] Weisse, A. K. and Bois, P., “Topographic effects on statistical characteristics of heavy rainfall and mapping in the
French Alps”, J. Applied Meteoro., 40, 720-740 (2001).
[11] Marquínez, J., Lastra, J. and García, P., “Estimation models for precipitation in mountainous regions: the use of GIS
and multivariate analysis”, J. Hydro., 270, 1-11 (2003).
[12] Hofierka, J., Parajka, J., Mitasova, M. and Mitas, L., “Multivariate interpolation of precipitation using regularized
spline with tension”, Transactions in GIS, 6, 135-150 (2002).
[13] Lloyd, C. D., “Assessing the effect of integrating elevation data into the estimation of monthly precipitation in Great
Britain”, J. Hydro., 308, 128-150 (2005).
[14] Sun, R. H., Zhang, B. P. and Tan, J., “A multivariate regression model predicting precipitation in the Daqing
mountains”, Mountain Res. and Develop., 28, 318-325 (2008), doi: 10.1659/mrd.0944.
[15] Odeh, I., McBratney, A. and Chittleborough, D., “Further results on prediction of soil properties from terrain
attributes: heterotopic cokriging and regression-kriging”, Geoderma, 67, 215-226 (1995).
[16] T. Hengl, G. Heuvelink, and A. Stein, “A generic framework for spatial prediction of soil variables based on
regression-kriging”, Geoderma, 122, 75-93 (2004).
[17] Hengl, T., Heuvelink, G. and Rossiter, D., “About regression-kriging: from equations to case studies”, Computers &
Geosciences, 33, 1301-1315 (2007).
[18] Webster, R. and Oliver, M., [Geostatistics for environmental scientists statistics in Practice], Wiley, Chichester,
p271 (2001).
[19] Goovaerts, P., Geostatistics for Natural Resources Evaluation, Oxford University Press, New York, p483 (1997).
[20] Nalder, I. A.and Wein, R. W., “Spatial interpolation of climatic normals: test of a new method in the Canadian
boreal forest”, Agri. and Forest Meteoro., 92, 211-225 (1998).
[21] Hill, R. C., Griffiths, W. E. and Judge, G. G., [Using EViews for undergraduate econometrics (2rd Edition)],
Dongbei University of Finance & Economics Press, Da Lian (2006).
[22] Gujarati, D. N., [Essentials of econometrics (3rd Edition)], McGraw-Hill, New York (2006).
[23] Liu, C., Qian, Z., Wu,,M., Song,,M. and Liu, J., “A composite study of the synoptic differences between major and
minor dust storm springs over the China-Mongolia areas”, Terres., Atmos. and Oceanic Sci., 15, 999-1018 (2004).
[24] J. Simpson, R. F. Adler, and G. R. North, “A proposed Tropical Rainfall Measuring Mission (TRMM) satellite”,
Bul. Ameri. Meteoro. Soc., 69, 278-295 (1988).
[25] Kummerow, C., Simpson, J., Thiele, O., Barnes, W., Chang, A. T. C. and Stocker, E., “The status of the Tropical
Rainfall Measuring Mission (TRMM) after two years in orbit”, J Appli. Meteoro., 39, 1965-1982 (2000).
[26] Immerzeel, W. W., Rutten, M. M. and Droogers, P., “Spatial downscaling of TRMM precipitation using vegetative
response on the Iberian Peninsula”, Re. Sen. Envir., doi: 10.1016/j.rse.2008.10.004 (2008).

Table 1. Descriptive statistics and Kolmogorov-Smirnov’s test for the annual period, the dry and the wet season: M-mean,
MIN-minimum, MAX-maximum, SD-standard deviation, K-S-Kolmogorov-Smirnov’s test value. The last seven
columns give the linear correlation coefficient between rainfall and predictors and their transforms.
Period Rainfall (mm) K-S Correlation
M MIN MAX SD X Y ELE SLP ASP PC1 PC2 PC3
Wet 297 221 410 39.4 0.75 -0.20 -0.70** 0.42** 0.55** 0.12 0.48** -0.66** 0.04
Dry 136 88.4 221 26.5 0.86 -0.21 -0.46** 0.23 0.36** -0.01 0.33* -0.42** -0.07
Annual 434 331 599 60.5 1.05 -0.22 -0.65** 0.37** 0.51** 0.07 0.46** -0.61** -0.00

** correlation is significant at the 0.01 level (2-tailed), * correlation is significant at the 0.05 level (2-tailed)

Proc. of SPIE Vol. 7492 74924I-6

 Table 2. Summary results of the step-wise regression analysis for precipitation in the Laohahe basin and improved
coefficient estimates (n = 44).

Regression Regression
Period Selected Predictors
coefficients (OLS†) coefficients (GLS‡)
intercept 417.6972 409.6250
PC1 26.89074 41.90602
Annual
PC2 -31.13085 -16.73457
PC22 16.15373 13.67581
intercept 285.9751 282.9170
PC1 18.89804 26.52168
Wet
PC2 -22.32165 -19.40965
PC22 11.50941 11.65545
intercept 136.2771 128.5539
Dry PC1 9.201015 16.36250
PC2 -8.373646 3.030121

†OLS-Ordinary least square estimation, ‡GLS-General least square estimation based on the weighted least square estimation of
residuals.

Table 3. Results of the Jarque-Bera, Lagrange-Multilier and Godlfeld-Quandt test for the OLS and GLS residuals in the
annual period, the wet and the dry season: Prob.-probability (F-statistic).

Jarque-Bera Lagrange-Multilier Godlfeld-Quandt

Residual Period test test test
F value Prob. F value Prob. F value Prob.
Wet 1.49 0.473 0.106 0.90 1.94 0.1351
OLS Dry 12.68 0.001 0.009 0.93 4.35 0.02
Annual 6.34 0.041 1.125 0.33 4.36 0.0086
Wet 19.09 0.000 0.61 0.54 0.58 0.6304
GLS Dry 12.18 0.002 0.89 0.35 2.29 0.11

Table 4. The regression model between the OLS residual variance and rainfall and predictor’s transforms: Prob.-probability
(F-statistic).

Period Rainfall PC1 PC2

Model Prob. Model Prob. Model Prob.
Wet Quadratic 0.217 Linear 0.236 Linear 0.048
Dry Quadratic 0.000 Linear 0.050 Quadratic 0.004
Annual Quadratic 0.000 Linear 0.015 Quadratic 0.021

Table 5. Comparison of interpolation methods for bias (MAE) and accuracy of the prediction (RMSE) at 8 validation points,
OK-ordinary kriging, MR-OLS multiple regression, RK-GLS regression kriging, MAE-mean absolute error (mm),
RMSE-root mean square errors (mm), RMSEr-relative prediction error (%).

Period Method MAE RMSE RMSEr Measured rainfall (mm)

Whole year OK 32.15 44.50 73.60
MR 30.89 37.66 62.27 433.59
RK 23.82 27.01 44.67
Wet season OK 21.77 26.05 66.11
MR 16.39 22.58 57.29 299.32
RK 13.48 17.02 43.20
Dry season OK 21.04 28.63 108.12
MR 18.18 22.82 86.17 134.27
RK 15.45 18.54 70.00

Proc. of SPIE Vol. 7492 74924I-7

Fig. 1. Location of the study area and the 52 rain gauges divided into interpolation (·) and validation (+) sets.

Fig. 2. Mean monthly rainfall of the stations over the period of January 1986-December 2005. Bars represent standard
deviation.

Proc. of SPIE Vol. 7492 74924I-8

 Fig.3. Digital elevation model

Fig.4. Comparison of prediction methods: ordinary kriging (a), OLS multiple regression (b), and regression-kriging (c) for
the annual period (_1), the wet (_2) and the dry season (_3).

Proc. of SPIE Vol. 7492 74924I-9

View publication stats