
Research Article


Effect of Pixel Size on the Areal Storm Pattern Analysis using Kriging 

A. Akbari,
A. Abu Samah
and
F. Othman



ABSTRACT

Several factors influence the interpolation accuracy of point source rainfall data during storms. Types of storm, network density and interpolation method are the most significant factors. Usually, random distributed point source data is converted into regular distributed grids, through the appropriate method. Kriging is a stochastic method that simulates the spatial surface using sample points, based on bestfit SemiVariogram (SV) models. This study assesses the effect of pixel size on the performance of Kriging interpolation using the Gaussian SV model. Twenty eight rainfall stations located at the upper part of the Klang River Basin (KRB), Malaysia are selected and three storm events are investigated. Simple Kriging interpolation is applied using different pixel size ranges from 50 to 3000 m. This study shows that pixel size is important issue when explaining the spatial pattern of rainfall. Optimal pixel size depends on the variance, minimum distance of pair points or rainfall network density and the visualization requirement. It is difficult to maintain a certain pixel size but based on the drawn result, it can be concluded that a pixel size at the range of 200 to 500 m is more appropriate for this region.







INTRODUCTION
Representing geospatial data in a raster data model requires grid cells definition.
The cell size is sensitive to map scale, computer processing power, positional
accuracy, point density, spatial autocorrelation structure and complexity of
terrain (Hengl, 2006). It also depends on the desired
quality of the output map. Pixel size and its impacts on the spatial analysis
particularly in rainfall runoff modeling are the focus of several studies. The
majority of research focuses on the influence of grid size of Digital Elevation
Models (DEMs) for the calculation of the topographic indices (Cai
and Wang, 2006) and watershed runoff simulation (Molnar
and Julien, 2000). Hengl (2006) showed that using
resolution coarser than the coarsest legible resolution means that we are either
not respecting the scale of work, positional accuracy, inspection density, size
of objects being mapped or the complexity of terrain. Hydrological modeling
employs rainfall data as an important variable. As stated by Earls
and Dixon (2007), the representation of rainfall data and its accuracy are
controlled by the spatial distribution of the rainfall stations and the spatial
interpolation methods used which may or may not reflect the reality. Several
factors may affect the accuracy of point source rainfall data interpolation
such as type of storm, network/gauge density and interpolation methods are important
factors. A number of studies on characterizing of rainfall has been carried
out in Klang River Basin (KRB) (Niemczynowicz, 1987;
Bacchi and Kottegoda, 1995). The results indicate that generally, the spatial
correlation of rainfall decreases with distance and different correlation structures
are observed during different rainfall events. The spatial extension of thunderstorms,
which create most floods, is limited and there are no routines to account for
this in design of rainfall (Desa and Niemczynowicz, 1996).
There are also errors attached to estimating localized rainfall due to the effects
of inadequate temporal resolution, inadequate spatial coverage or network configuration,
inadequate gauge density and instrument error (PetersLidard
and Wood, 1994).
KRIGING THEORY AND APPLICATIONS
Kriging is named after D.G. Krige, a South African mining engineer and pioneer
in the application of statistical techniques to mine evaluation. The Kriging
technique is derived from the theory of regionalized variables which is the
invent of Matheron (1971). This theory is formed based
on the observation that the variabilities of all regionalized variables have
a particular structure (Journel and Huijbregts, 1978).
As an example, a storm event can be characterized by the spatial distribution
of a certain number of measurements. More details on the geostatistical theory,
Kriging and its algorithm can be found in (Matheron, 1971;
Journel and Huijbregts, 1978; Cressie,
1993; Isaaks and Srivastava, 1990; Webster
and Oliver, 2007). Here, some important elements of this theory are presented,
according to the Journel and Huijbregts (1978):
Regionalized variable (ReV): When a variable is distributed in space,
it is said to be ReV and the phnomeno that the ReV is used to represent is called
regionalization (Journel and Huijbregts, 1978). This definition
can be extended to almost; all georeferenced data or spatial variables such
as population density, canopy density, water quality, air pollution and so on.
In mathematic language, a ReV is a function f (x) which takes a value at every
point x, in space. ReV takes two contradictory characteristics, which are a
Random Variable (RV) and a certain functional representation. A RV is a variable,
which takes a certain number of numerical values according to a certain probability
distribution (Journel and Huijbregts, 1978).
Random function (RF): Random function expresses the random and structured aspects of a ReV. These aspects are as follows: • 
In each point x_{1},Z(x) is the RV 
• 
Z(x) is also a RF in the sense that for each pair of points
x_{1} and x_{1}+h, the corresponding RFs Z(x_{1})
and Z(x_{1}+h) are not, in general, dependent but are related by
a correlation expressing the spatial structure of the initial ReV Z(x) 
Mathematical expectation (first order moment): Consider a random function Z(x) at point x. if the distribution function of Z(x) has an expectation then, this expectation is generally a function of x and is written as:
Second order moments: The three secondorder moments considered in geostatistic
are as follows:
• 
The variance of Z(x). when this variance exists, it is defined
as the secondorder moment about the expectation m (x) of the RV Z(x), i.e., 
As with the expectation m(x), the variance is generally a function of x. • 
The covariance. It can be shown that if the two RVs Z(x_{1})
and Z(x_{2}) have variances at the point x_{1} and x_{2},
then they also have a covariance which is a function of two locations x_{1}
and x_{2} and is written as: 
• 
The variogram. The variogram function is defined as the variance
of the increment [Z(x_{1})Z(x_{2})] and is written as: 
The function γ( x_{1} , x_{2} ) is then the semivariogram (SV). When the RF is stationary of order 2, mathematical expectation exists and does not depend on the support point x, thus, E {Z(x)} = m. For each pair of RV {Z(x_{1}), Z(x_{1}+h)} the covariance exists and depends on the separation distance h. according to the Eq. 2 and 3 can be drive as follow: where, h represent vector of coordinates. The stationarity of the covariance implies the stationarity of the variance and the variogram. The following relations are immediately evident: The Eq. 4 indicate that under the hypothesis of secondorder stationarity, the covariance and the variogram are two equal tools for characterizing the autocorrelations between two variable Z(x+h) and Z(x) separated by a distance h.
Covariance and correlation are measures of the similarity between two different
variables, where the data pairs represent measurements of the same variable
made some distance apart from each other. The separation distance is usually
referred to as lag. The correlation versus lag is referred to as the correlogram
and the semivariance versus lag is the SV. Kriging employs SV model (Bohling,
2005).
In GIS environment, Kriging can be seen as a point interpolation, which requires
a point map as input and returns a raster map with estimations and optionally
an error map. The estimations are weighted averaged input point values. The
weight factors in Kriging are determined by using a userspecified SV model.
The estimated values are thus a linear combination of the input values and have
a minimum estimation error. The optional error map contains the standard errors
of the estimates.
Characteristics of the typical SV are shown in Fig. 1. Sill is the semivariance value at which the variogram levels off. Range is the lag distance at which the reaches the sill value. Presumably, autocorrelation is essentially zero beyond the range. Nugget is, in theory, the value at the origin (0 lag) should be zero. If it is significantly different from zero for lags very close to zero, then this value is referred to as the nugget, which represents variability at distances smaller than the typical sample spacing, including measurement error. For the sake of it is necessary to replace the empirical SV with an acceptable SV model. Part of the reason for this is that the Kriging algorithm will need access to SV values for lag distances other than those used in the empirical SV. More importantly, the SV models used in the Kriging process need to obey certain numerical properties in order for the Kriging equations to be solvable. Most frequently used models are spherical, exponential, Gaussian and power model.
Barancourt et al. (1992) have provided a comparison
of rainfall assessments by a global Kriging and geostatistical model. The prediction
performance is compared by using cross validation. It is found that ordinary
Kriging yields more accurate predictions than linear regression when the correlation
between rainfall and elevation is reasonable. Karamouz and
Araghinejad (2005) applied the Kriging method to evaluate monthly regional
rainfall in the central part of Iran. Thavorntam et al.
(2007) studied spatial interpolation of mean annual rainfall in a 29 year
period using interpolation methods such as Inverse Distance Weighting (IDW),
radial basis function and ordinary Kriging.
 Fig. 1: 
Characteristics of the typical SVSemivariogram 
They indicated that ordinary Kriging
with IDW a spherical model has better performance for interpolation of rainfall
within the region of Thailand. However, Dirks et al.
(1998) studied interpolation of rainfall data on the Norfolk Island,
Australia. Thirteen rain gauges on the area of 35 km^{2} are used. They
found that Kriging did not provide a significant improvement over simple methods
like the IDW, the Thiessen and the area mean methods because of the demanding
on computations.
MATERIALS AND METHOD
Study area: This study is conducted at University Malaya as a part of
postgraduate research project within 6 month started from November, 2008. The
study is carried out on the upper parts of the Klang River Basin located in
Kuala Lumpur/Malaysia. The Klang Valley is located between 10l° 30’
10l° 55’ longitude and 3° 3° 30’ latitude. It falls largely
within the state of Selangor and encompasses the entire Federal Territory of
Kuala Lumpur (Fig. 2). The basin area at the outlet is about
675 km^{2}. The mean annual rainfall based on the Association of Southeast
Asian Nations Climate Atlas for Peninsular Malaysia is about 2400 mm (Tick
and Samah, 2004). The elevation above sea level ranges from 20 m at the
outlet to 1420 m upstream.
Data set used: Rainfall data for 28 stations located at the upper Klang river basin are collected form Department of Irrigation and Drainage (DID) of Malaysia. General characteristics of the rainfall stations such as ID that is the key code indicated by DID and geographic coordinates of stations are provided in Table 1.
Software used: The Integrated Land and Water Information System (ILWIS)
which is free rasterGIS software, developed by International Institute for
GeoInformation Science and Earth Observation (ITC), The Netherlands (Koolhoven
et al., 2007) is used for this study.
Table 1: 
General information of rainfall stations used in this study 

 Fig. 2: 
Study area and rainfall stations network at upper Klang River
Basin 
Available at http://www.itc.nl/ilwis/default.asp.
RESULTS AND DISCUSSION
The three random rainfall events occurred on May 6, Sep. 6 and Dec. 21 in the
year 2002, are selected. General statistics such as minimum, maximum, average,
standard derivation and variance are calculated for the selected events (Table
2). A point map is generated based on available rainfall stations shown
in Fig. 2. Total recorded rainfall in each station for that
particular date is then linked to the point map as attribute. Spatial autocorrelation
(Odland, 1988) applied to estimate the Kriging parameters
included nugget, sill and range effects (Borgan and Vizzaccaro,
1997). As shown in Fig. 3, Gaussian SV model fits through
the experimental SV, resulted from spatial autocorrelation as better model compared
to the exponential and spherical model (Akbari et al.,
2008). Simple Kriging is used, because no certain direction can be found
in the storm propagation in this region (Desa and Niemczynowicz,
1996). Kriging interpolation technique is then applied to the selected storms.
Different Pixel size range from 50 to 3000 m is examined for all storms. Kriging
interpolation using point map in ILWIS provides raster map with predefined pixel
size.
 Fig. 3: 
Experimental and empirical SVs for the event of 21 December,
2002 
Table 2: 
Statistical summaries of investigated storm events 

A histogram is made for the raster map including pixel values and the
corresponding number of pixels, the fraction of total pixels, the area and statistical
summaries including minimum, maximum, average, standard deviation and summation
of cell values. Table 3 is provided as an example for derived statistical
summaries of rainfall event of Dec 21, 2002. This table does not provide for
the two o ther events. However, a comparison is presented to show the variation
of pixel size with the estimated average rainfall (Fig. 3,
4), standard deviation (Fig. 5), area allocated
by average rainfall (Fig. 6).
Table 3: 
Statistical summaries of raster map showing Kriging interpolation
based on different pixel size for the Storm Dec 21 2002 

A: Pixel size (m), B: Statistical terms, C: Total rainfall
during the storm (mm), D: The number of pixels with a certain value or meaning
(pixel frequencies), E: The cumulative percentage of pixels with this value
or a smaller value (%), F: The relative cumulative numbers of pixels with
this value or a smaller value as a percentage of the total number of pixels
in the map, G: The area of pixels with a certain value or meaning as D*
pixel size *pixel size (m^{2}) 
 Fig. 4: 
Variations of average rainfall vs. pixel size 
 Fig. 5: 
Standard deviation of cell values vs. pixel size 
One can see the variation of
pixel size with the total number of pixels assigned by maximum value in that
particular pixel size in Fig. 7 and file size, which is important
in computer processing and allocating memory in Fig. 8.
 Fig. 6: 
Variation of area representing the average rainfall vs. pixel
size 
 Fig. 7: 
Variation of number of pixel assigned by maximum value vs.
pixel size 
Finally, we examined the effect of pixel size on the visualization, which is
shown in Fig. 9a and b and 10.
These maps are based on the storm event of Dec 21, 2002.
 Fig. 8: 
Variation of file size vs. pixel size 
Rainfall distribution in space can be explained by the standard deviation and
variance of samples. Statistical analysis for the recorded rainfall at the rainfall
stations indicates the high variance for the event of Sept. 6, 2002 compare
to the 2 other investigated events. Based on the results of the interpolations,
it is possible to interpret the spatial variation of rainfall based on the investigated
pixel size and its effect on the statistical terms such as average, standard
deviation and area related to the maximum value in each map. As shown in Fig.
4, the average rainfall is sensitive to the pixel sizes at about 50750
m. Almost in three tested storms, average rainfall does not change by taking
the pixel size larger than 750 m. Based on the variance of the events; different
trends can be seen for each event. For example, for the event of Sept. 6, 2002
that is the highest variance (945) among the samples, average value is more
sensitive to the pixel size. However, for the storm event on the Dec. 21, 2002
that has the lowest variance, variation of the average value is not sensitive
to the pixel size. That means the enlargement of grid resolution leads to aggregation
or up scaling and the decrease of grid resolution leads to disaggregation or
down scaling (Hengl, 2006). As the grid becomes coarser,
the overall information content in the map decreases progressively and vice
versa (Hengl, 2006). Figure 5 shows
almost the same trend for the variaton of pixle size with standard deviation.
The trends shown in this figure imply on the sensitivity of the standard deviation
to the pixel size from 50 to 750 m. The number of pixels in the raster map also
expresses the surface values. Figure 6 shows the variation
of pixel size in relation to the area coreponding to the mean value for three
storms. It can be seen that the area related to the mean value increases by
selecting coarser pixel size at an exponential rate. This is due to the smoothing
effect of a mean value increases by selecting coarser pixel size at an exponential
rate. This is due to the smoothing effect of a larger pixel size. The smoothing
effect is confirmed by Fig. 7, which illustrates the variation
of the number of pixeles assigned by the maximum value for different pixel size
in three investigated events. High sensitivity can be found for the pixel sizes
at about 50 to 500 m. Figure 8 shows clearly the effect of
the pixel size on the size of file.
 Fig. 9: 
Storm distribution based on the Kriging interpolation for
different pixel size (a) (1250 m) and (b) (50 m) 
This become important issue when working with database and transfering and
analyzing raster data. Selection of the optimal pixel size facilitates the easier
transfer and processing.
 Fig. 10: 
Isorain lines derived from the Kriging interpolation for different
pixel size (1250 m), left, (50 m), right 
Eventually, a comparison is made to show the performance
of pixel size on the spatial distribution of rainfall derived from kriging interpolation
(Fig. 9). These two maps emphesize the effect of pixel size on the visualization especially
when isoline is derived (Fig. 10).
CONCLUSION This study shows clearly that pixel size has a significant influence on storm pattern analysis. Variance of the rainfall over the area has a meaningful effect on the selection of the legible pixel size and therefore, on the result of points interpolation. In fact, variance is comparable with the complexity of terrain when Digital Elevation Model (DEM) is generated from the points or contour lines. It can be concluded that the rainfall variance over the space is important factor in selecting appropriate pixel size when using Kriging interpolation. High variance leads to take the small pixel sizes and low variance leads to take larger pixel size. It is because of the fact that small pixel size is more able to explain the rainfall variation. No ideal pixel size exists, but rather a range of suitable pixel size. Optimal pixel size will depend on the variance, minimum distance and the visualization requirement. It is recommended to select a smaller pixel size when the variance of the source point values is high. Small pixel size provide better visualization but generates larger file size. Raster maps in Fig. 9, results from Kriging interpolation for the rainfall event of 21 Dec. 2002. As shown in Fig. 10, small grid size is more advisable when contour line drawing is required from the raster map. Hence, one can report that variation of average rainfall to the pixel size illasterates sensitivity of the pixel size to the rainfall variance and recommended range of pixel size at 200 to 500 m scale in this region. Pixel size has a cosiderble effect on the visualization and the file size.
ACKNOWLEDGMENTS Financial support of this study is provided by University of Malaya through the research fund project number PS017/2007B. Rainfall data provided by DID Malaysia. We would like to appreciate to all DID staffs of hydrologic section for kindly communication. We also appreciate ITC for providing free GIS software. Valuable contribution of Professor Azizan Abu Samah is appreciated.

REFERENCES 
1: Akbari, A., B.F. Othman and A.S. Azizan, 2008. Spatial storm pattern analysis using Kriging, case study: Upper Klang River Basin. Malaysian Science and Technology Congress (MSTC2008). December 1617 2008, KLCC, Kuala Lumpur, Malaysia.
2: Bacchi, B. and A.N.T. Kottegoda, 1995. Identification and calibration of spatial correlation patterns of rainfall. J. Hydrol., 165: 311348. CrossRef  Direct Link 
3: Barancourt, C., J.D. Creutin and J. Rivoirard, 1992. A method for delineating and estimating rainfall fields. Water Resour. Res., 28: 11331144. Direct Link 
4: Borgan, M. and A. Vizzaccara, 1997. On the interpolation of hydrologic variables: Formal equivalence of multi quadratic surface fitting and kriging. J. Hydrol., 195: 160171. Direct Link 
5: Desa, M.N. and J. Niemczynowicz, 1996. Spatial variability of rainfall in Kuala Lumpur, Malaysia: Long and short term characteristics. Hydrol. Sci. J., 41: 345362. Direct Link 
6: Dirks, K.N., J.E. Hay, C.D. Stow and D. Harris, 1998. Highresolution studies of rainfall on Norfolk Island: Part II: Interpolation of rainfall data. J. Hydrol., 208: 187193. CrossRef  Direct Link 
7: Earls, J. and B. Dixon, 2007. Spatial interpolation of rainfall data using ArcGIS: Comparative study. ESRI User Conference, 31Oct2007, San Diego, California, USA.
8: Hengl, T., 2006. Finding the right pixel size. Comput. Geosci., 32: 12831298. CrossRef  Direct Link 
9: Karamouz, M. and S. Araghinejad, 2005. Advanced Hydrology. 1st Edn., Amirkabir University Press, Tehran, Iran, ISBN: 94646327405
10: Koolhoven, W., J. Hendrikse, W. Nieuwenhuis, B. Retsios and M. Schouwenburg et al., 2007. The Integrated Land and Water Information System. Enschede, The Netherlands
11: Molnar, D.K. and P.Y. Julien, 2000. Gridsize effects on surface runoff modeling. J. Hydrol. Eng., 5: 816. Direct Link 
12: Niemczynowicz, J., 1987. Storm tracking using rain gauge data. J. Hydrol., 93: 135152. Direct Link 
13: Obled, C., J. Wendling and K. Beven, 1994. The sensitivity of hydrological models to spatial rainfall patterns: An evaluation using observed data. J. Hydrol., 159: 305333. Direct Link 
14: PetersLidard, C.D. and E. Wood, 1994. Estimating storm areal average rainfall intensity in field experiments. Water Resour. Res., 30: 21192131. Direct Link 
15: Thavorntam, W., N. Wattanakij and C. Mongkolsawat, 2007. Estimation of rainfall spatial interpolation methods in NorthEast Thailand. Asian Conference on Remote Sensing (ACRS 2007) (PWTC), November 1216 2007, Kuala Lumpur, Malaysia.
16: Tick, L.J. and A.A. Samah, 2004. Weather and Climate of Malaysia. 1st Edn., University Malaya Perss, Kuala Lumpur, ISBN: 9831001761
17: Bohling, G., 2005. Kriging. Kansas Geological Survey. http://people.ku.edu/~gbohling/cpe940/Kriging.pdf.
18: Cai, X. and D. Wang, 2006. Spatial autocorrelation of topographic index in catchments. J. Hydrol., 328: 581591. CrossRef  Direct Link 
19: Cressie, N.A.C., 1993. Statistic for Spatial Data. John Wiley and Sons, New York, ISBN10: 0471002550, pp: 928
20: Isaaks, E.H. and R.M. Srivastava, 1990. An Introudction to Applied Geostatistics. Oxford University Press, New York, USA., ISBN10: 0195050134, pp: 592
21: Journel, A.G. and C.J. Huijbregts, 1978. Mining Geostatistics. 1st Edn., Academic Press, London, ISBN10: 0123910501, Pages: 600
22: Matheron, G., 1971. The theory of regionalized variables and their applications. Cahiers du Centre de Morphologie Mathématique, Centre de Geostatistique, Paris, pp: 212.
23: Odland, J., 1988. Spatial Autocorrelation. Sage University Publications, USA., ISBN10: 0803926529, pp: 88
24: Webster, R. and M.A. Oliver, 2007. Geostatistics for Environmental Scientists. 2nd Edn., John Wiley and Sons, UK., ISBN13: 9780470209394



