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

net/publication/249604761

Sensitivity Analysis of Quality Assurance Using the Spatial Regression

Approach---A Case Study of the Maximum/Minimum Air Temperature

Article  in  Journal of Atmospheric and Oceanic Technology · October 2005

DOI: 10.1175/JTECH1790.1

CITATIONS READS

51 99

2 authors:

Kenneth G. Hubbard Jinsheng You

University of Nebraska at Lincoln University of Nebraska at Lincoln
221 PUBLICATIONS   6,340 CITATIONS    34 PUBLICATIONS   517 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jinsheng You on 22 July 2016.

The user has requested enhancement of the downloaded file.

1520 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

Sensitivity Analysis of Quality Assurance Using the Spatial Regression Approach—

A Case Study of the Maximum/Minimum Air Temperature
KENNETH G. HUBBARD AND JINSHENG YOU
High Plains Regional Climate Center, University of Nebraska at Lincoln, Lincoln, Nebraska

(Manuscript received 8 October 2004, in final form 14 March 2005)

ABSTRACT

Both the spatial regression test (SRT) and inverse distance weighting (IDW) methods have been applied
to provide estimates for the maximum air temperature (Tmax) and the minimum air temperature (Tmin) in
the Applied Climate Information System (ACIS). This is critical to the processes of estimating missing data
and identifying suspect data and is undertaken here to ensure quality data in ACIS. The SRT method was
previously found to be superior to the IDW method; however, the sensitivity of the performance of both
methods to input parameters has not been evaluated. A set of analyses is presented for both methods
whereby the sensitivity to the radius of inclusion, the regression time window, the regression time offset, and
the number of stations used to make the estimates are examined. Comparisons were also conducted
between the SRT and the IDW methods. The performance of the SRT method stabilized when 10 or more
stations were applied in the estimates. The optimal number of stations for the IDW method varies from only
a few to 30. The results indicate that the best estimates obtained using the IDW method are still inferior to
the worst estimates obtained using the SRT method.

1. Introduction designed for the single station review of data to detect

The spatial regression test (SRT) method has been potential outliers (Wade 1987; Meek and Hatfield 1994;
found to be superior to the inverse distance weighting Eischeid et al. 1995).
(IDW) method (You et al. 2004, manuscript submitted Recently, the use of multiple stations in quality as-
to J. Atmos. Oceanic Technol., hereafter YHG) when surance procedures has proven useful; for example the
applied to provide estimates for the maximum air tem- spatial tests compare a station’s data against the data
perature (Tmax) and the minimum air temperature from neighboring stations (Wade 1987; Gandin 1988;
(Tmin) in the Applied Climate Information System Eischeid et al. 1995; Hubbard 2001a). The spatial tests
(ACIS). However, the sensitivity of the performance of involve the use of neighboring stations to make an es-
both methods to the input parameters has not been timate of the measurement at the station of interest.
evaluated. This paper conducts a sensitivity analysis on The IDW technique weights the values at surrounding
the performance of the SRT and IDW methods of es- stations according to the inverse of the distance sepa-
timating missing data. We examine the effect of dis- rating the locations (Guttman et al. 1988; Wade 1987),
tance to the surrounding stations and the number of while other statistical approaches seek to provide a
surrounding stations on overall performance. nonbiased estimate [e.g., multiple regression (Eischeid
Quality assurance (QA) procedures have been ap- et al. 1995, 2000) and the bivariate linear regression test
plied (Guttman and Quayle 1990) to (semi) automati- (Hubbard et al. 2005)].
cally check the validity of weather data from the coop- Unlike the IDW, the spatial regression test used
erative climatological stations in the National Climatic herein (Hubbard et al. 2005) does not assign the largest
Data Center (NCDC). General testing approaches, weight to the nearest neighbor but, instead, assigns the
such as the threshold method and step change test were weight according to the root-mean-square error (rmse)
between the station of interest and each of the neigh-
boring stations. Research has demonstrated the excel-
Corresponding author address: Kenneth G. Hubbard, High
Plains Regional Climate Center, University of Nebraska at Lin- lent performance of the spatial regression test in iden-
coln, Lincoln, NE 68583-0728. tifying seeded errors (Hubbard et al. 2005). In a sepa-
E-mail: [email protected] rate study, the investigators used the spatial regres-

© 2005 American Meteorological Society

JTECH1790
OCTOBER 2005 HUBBARD AND YOU 1521

sion test to identify the potential outliers during unique density of stations also varies significantly within these
weather events. In the case of hurricanes, cold front four states. Based on results from a previous study
passage, floods, and droughts, the number of QA fail- (YHG), we concluded that these states represent the
ures was largely due to the different times of observa- range of conditions over which we wished to test the
tion coupled with the ambiguity associated with posi- sensitivity of the methods. California was chosen be-
tion relative to tight gradients of temperature or pre- cause the coastal stations there were found to have rela-
cipitation. tively poor performance for both the SRT method and
The spatial regression approach was found superior IDW method compared to the case of the plains sta-
to the inverse distance approach for the maximum air tions (YHG). Nevada was selected for relatively poor
temperature (Tmax) and the minimum air temperature performance in remote parts of the state due to sparsely
(Tmin) in a previous study (YHG), with the largest im- distributed stations in mountainous regions. The per-
provements in the coastal and mountainous regions. formance of SRT and IDW was better in Utah, which
Both methods were found to perform relatively poorer was also selected for sensitivity analysis because it has
when the weather stations were sparsely distributed mountainous regions, like Nevada, but generally higher
(YHG). The success of the spatial regression approach station density. Nebraska was chosen as an example for
is in part due to its implicit ability to resolve the sys- the plains states, with adequate station density, except
tematic differences caused by temperature lapse rate in the Sand Hills region in the north-central portion of
with elevation, which is not accounted for in the inverse the state.
distance weighting method. The data from stations in the four states within the
The spatial regression and inverse distance weighting cooperative observer weather data network, a regional
automated weather data network (Hubbard 2001b),
methods provide separate estimates of a station’s data
and other networks were retrieved through the ACIS, a
based on surrounding stations. This is critical to the
distributed data management system. This study in-
processes of identifying suspect data and is undertaken
cludes estimation of the daily maximum (Tmax) and
here to ensure quality data in the ACIS (Hubbard et al.
minimum air temperature (Tmin) for the period.
2004). The estimates are also used to form a continuous
dataset by filling in missing values. The spatial regres-
sion test has three parameters that the user may adjust 3. Methods
to obtain estimates. These are the station list, the width
a. Spatial regression test
of the time window, and the time offset related to the
window. The inverse distance weighting technique has The spatial regression test (Hubbard et al. 2005) is a
one parameter, the number of stations used in the quality control approach that checks whether the vari-
weighted estimate. In the previous study the number of able falls inside the confidence interval formed from
stations used in the IDW method was five. The radius surrounding station data during a time period of length
of inclusion for the SRT method was set to 50 km, (window length) n. All stations (M ) within a certain
except in those areas where no station was located distance of the station of interest are selected, and a
within 50 km. In such cases the radius of inclusion was linear regression performed for each station paired with
increased to 150 km. This paper conducts a sensitivity the station of interest and centered on the datum of
analysis of the performance of both the spatial regres- interest. For each surrounding station, a regression-
sion test and inverse distance weighted method for both based estimate (xi ⫽ ai ⫹ biyi) is formed. The weighted
Tmax and Tmin to the distance to the reference stations estimate (x⬘) is obtained by utilizing the standard error
and the number of reference stations. The performance of estimate (s), also known as rmse in the weighting
of each estimation approach was evaluated, and the process:
optimum parameters are suggested.
x⬘ ⫽ ⫾ 冉兺
N

i⫽1
⫾ xi2si⫺2 冊 冉兺 冊
0.5 N

i⫽1
si⫺2
⫺0.5
, 共1兲

2. Data where N is the number of stations to be used in the

estimate (generally restricted to an R2 greater than 0.5
The tests were carried out over four states, California within a given radius of inclusion, e.g., 50 km). Note
(CA), Utah (UT), Nevada (NV), and Nebraska (NE), that N is selected by the user and N ⱕ M. Care must be
for the year 2000. These four states represent the di- taken to preserve the correct sign on x⬘. To account for
verse landforms and topographical features from possible systematic time shifting of observations (this
coastal regions to mountainous regions to plains. The occurs when an observer consistently writes his obser-
1522 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

vation down on the day before or after the actual date get station by their proximity to the target station. This
of observation), the surrounding station’s data are each procedure is given by

冒兺
shifted by ⫾1 day and the regression repeated. The n n
time shift (⫺1, 0, or ⫹1) that results in the lowest stan-
dard error of estimate is then taken into (1) and (2).
x̂ ⫽ 兺
i⫽1
共xiwi兲
i⫽1
wi, 共4兲

The weighted standard error of estimate (s⬘) is calcu- where x̂ is the IDW estimate, xi is the particular mea-
lated: surement at the ith surrounding station, and the weight
N function wi is derived from the inverse of the distance
兺
1 between the target station and the ith surrounding sta-
⫽ N⫺1 si⫺2. 共2兲
s⬘2 i⫽1 tion. In this study, the number of stations (n) was varied
to create different sets of estimates, and the perfor-
Now the confidence intervals are formed from s⬘, and mance of each set relative to the original measurements
we test whether or not the station value (x) falls within was determined. Many simple weighting methods have
the confidence intervals: been applied for distance-based methods such as the
inverse square distance weighting and the exponential
x⬘ ⫺ fs⬘ ⱕ x ⱕ x⬘ ⫹ fs⬘. 共3兲
distance weighting methods. In this paper we examine
If the relation in Eq. (3) holds, then the corresponding the simple IDW method.
datum passes the spatial regression test. Increasing f
decreases the number of potential Type I errors but c. Evaluation of estimations
increases the number of potential Type II errors. Un- The Nash and Sutcliffe coefficient of efficiency
like distance weighting techniques, we can sort the rmse (NSC) (Nash and Sutcliffe 1970) was used as a measure
into ascending order and select those stations that com- of fit between actual and estimated data. The quantity
pare most favorably to the station of interest. By re- NSC is given by

冋兺 册冒 兺
peating the process for different values of N we can m m m
observe how the performance of the method changes
with N. These may or may not be the closest stations.
NSC ⫽
i⫽1
共xi ⫺ x兲2 ⫺ 兺 共x̂ ⫺ x 兲
i⫽1
i i
2

i⫽1
共xi ⫺ x兲2,

The spatial regression test has parameter settings

共5兲
(window length, offset, and station list) that are speci-
fied for the quality assurance processes. Window length where xi is the measured variable, x̂i is estimated vari-
is the number of data pairs applied to form a regression. able, and x៮ i is the arithmetic mean of xi for all events
Offset determines where the window is located relative i ⫽ 1 to m.
to the datum to be checked. For example, in our analy- The calculation of NSC is a procedure that essentially
ses an offset of zero centers the widow on the datum of sums the deviations of the observations from a linear
interest, so aligning the datum of interest with the be- regression line with a slope of 1. If the measured vari-
ginning of the window would be specified by an offset able is estimated exactly for all observations, the value
of ⫺29 for a 60-day window. On the other hand, align- of NSC is 1. Low values of NSC show high deviations
ing the datum of interest with the end of the window between measured and estimated values. If NSC is
would be accomplished by setting the offset to 30 for a negative, estimates are very poor and population aver-
60-day window. The station list contains the stations age values would provide better estimates.
used in estimating the data. The station list can be Values of R2, the explained variance, have been
sorted or unsorted stations obtained using different ap- widely used as an index of agreement, so R2 is included
proaches, for example, stations located within a certain here also. However, R2 is not always instructive and, as
distance of the station of interest or within a rectangu- Willmott (1981) cautions, should not be used alone to
lar neighborhood centered on the station of interest. In assess the accuracy of estimates. In this paper, Will-
this study, we determine the neighborhood by specify- mott’s (1981) D, the index of agreement for assessing
ing the radius of inclusion where all stations (M ) within model performance is also used and is expressed as
the resulting circle are considered for the station list.

b. Inverse distance weighting method

D⫽1⫺ 兺
n

i⫽1
共x̂i ⫺ xi兲2 冒兺 n

i⫽1
共ⱍx̂i ⫺ xⱍ ⫹ ⱍxi ⫺ xⱍ兲2.

The IDW method is a simple distance weighted esti- 共6兲

mate of the value at the target station. The assumption The D index and NSC are more sensitive to system-
here is that surrounding stations are related to the tar- atic model error than are R and R2 and reflect system-
OCTOBER 2005 HUBBARD AND YOU 1523

atic model bias when coupled with the R2 statistic. Val- no sorting on the rmse so the performance of SRT is
ues of D range from 0.0 for complete disagreement to not optimal. One example is shown in Figs. 1a and 1b.
1.0 for perfect agreement. Other measures of model Patterns vary somewhat from station to station; how-
performance included in this paper are the systematic ever, when the radius is greater than 80 km, D, NSC,
(Es) and nonsystematic (Eu) components of the rmse: R2, Es, Eu, and rmse for the SRT method are relatively

冋兺 冒册
m
stable or decline only slightly. When the radius is less
0.5
Es ⫽ 共Pri ⫺ xi兲2 m , 共7兲 than 80 km, indices at some stations have relatively
i⫽1 large fluctuations.

冋兺 冒册
m
An example of the analysis of SRT method for sen-
0.5
Eu ⫽ sitivity to window length is shown for the window (Figs.
共x̂i ⫺ Pri兲2 m , 共8兲
i⫽1 1c,d) and the window offset (Figs. 1e,f), respectively. In
general, D, NSC, R2, Es, Eu, and rmse diverge for win-
rmse ⫽ 关共Es2 ⫹ Eu2 兲兴0.5, 共9兲 dows from 20 to 150 days (the largest window length
where Pri is calculated from the slope (b) and intercept evaluated in this study). When more stations are exam-
(a) of the regression of estimated x and observed x ined, the estimates obtained are relatively stable when
(such that Pri ⫽ a ⫹ b xi). the window length is larger than 60 days. When the
window length is less than 60, the fluctuations of D,
NSC, R2, Es, Eu, and rmse are larger. The performance
4. Results of SRT changes slightly with the window offsets. In
The estimates using the spatial regression test were general, data estimates are best when the window is
evaluated against the measurements for stations within centered on the datum of interest (time shift equal to
the four states (California, Nevada, Utah, and Ne- zero). Figures 1e and 1f demonstrate that using the time
braska) for the year 2000. The settings used for the shift of ⫺1 is only slightly different from using a time
sensitivity analysis of the spatial regression test are shift of ⫹1. We suggest using a window length larger
shown in Table 1. For example, in the radius analysis, than or equal to 60 days and using the centered offset
the length of the time window and the time offset were (offset being half of the window) for best results.
set to 60 and 0, respectively, while the radius varied This study evaluated the performance of the SRT
from 16 to 241 km. For each radius, all available sta- under a second implementation according to the num-
tions within the circle were applied during the analysis. ber of stations (see Figs. 2 and 3). In this analysis, the
For the analysis on number of stations in the SRT reference stations were sorted based on the rmse be-
method, the stations within 160 km of the station of tween the daily time series of 2000 at the current station
interest were ranked based on the rmse obtained from and the daily data of 2000 at the reference stations.
regression between the measurements at reference sta- When less than 10 stations were applied in the estima-
tions and measurements at the station of interest. For tion, the performance of the SRT method in some cases
the IDW method, the stations were sorted by distance varied significantly with the number of stations, for ex-
to the station of interest. We illustrate the findings by ample, those shown in Fig. 2. All indices stabilized
showing the results from individual stations selected when 10 or more stations were used to estimate the
from different topographic regions: coastal, plains, Tmax or Tmin. In Fig. 3 we show the proportion of sta-
mountains, and basins. tions (P) that have a value of NSC, R2, and rmse
The SRT method was initially analyzed based on the smaller than the value on the x axis for all stations in
inclusion of all stations within the radius of inclusion, four states. The distribution curve of these indices for
which ranged from 16 to 241 km. In this case, there was estimates obtained using 10 stations is close to that us-

TABLE 1. Settings for the sensitivity analysis.

Settings during the sensitivity analysis

Sensitivity study conducted on Window length Offset Radius No. of stations
Window length Variable Middle of series 80 km All stations within 80 km
Time offset 60 days Variable 80 km All stations within 80 km
Radius 60 days Middle of series Variable All stations within radius
No. of stations 60 days Middle of series 160 km Variable sorted stations within 160 km
1524 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

FIG. 1. Examples of performance of the SRT method with regard to (top) distance, (middle) window length,
and (bottom) offset for three different locations: San Luis, CA; Minden, NE; and Fish Springs, UT.

ing more stations. The differences of rmse at 90% prob- The performance of the IDW method was also evalu-
ability for 10 stations and 30 stations for both Tmax and ated to determine its sensitivity to the number of sta-
Tmin are less than 0.5°F, which is within the expected tions included (Figs. 4 and 5). The reference stations
observational errors (⫾1°F). In this study, 10 stations are sorted by the distance to the station of interest. As
are recommended for the estimations of Tmax and Tmin shown in the examples, D, NSC, R2, Es, Eu, and rmse
using SRT method. vary with the number of stations and differ for different
OCTOBER 2005 HUBBARD AND YOU 1525
1526 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

FIG. 3. The proportion of stations (P) in the four states (CA, NE, NV, UT) that have a value of (top) NSC, (middle) R2, or
(bottom) rmse smaller than the value shown for the SRT method.

stations. No specific number of stations shows stability port, CA, stations have better estimates due to a higher
of these indices (Fig. 4). Table 2 lists the number of station density. In contrast, the sparse distribution
stations falling within the circle of 32, 80, and 160 km. around Merced, CA, and Partoun, UT, results in rela-
The Minden, NE, and San Francisco International Air- tively poor estimates. The sensitivity to the number of

←
FIG. 2. Examples of performance of the SRT method with regard to number of stations
sorted by least rmse between the observations at station of interest and observations of
reference stations: (top to bottom) San Francisco International Airport, CA; Merced, CA;
Minden, NE; and Partoun, UT.
OCTOBER 2005 HUBBARD AND YOU 1527
1528 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

FIG. 5. The proportion of stations (P) in the four states (CA, NE, NV, UT) that have a value of (top) NSC, (middle) R2, or
(bottom) rmse smaller than the value shown for the IDW method.

stations also changed for the densely distributed sta- using five stations is recommended for the IDW
tions and sparsely distributed stations. In general, the method for both Tmax and Tmin.
distributions of NSC, R2, and rmse using five stations The distribution of NSC, R2, and rmse for the SRT
are close to those using more stations (see Fig. 5). Thus, method and the patterns for single stations demonstrate

←
FIG. 4. Examples of performance of the IDW method for a given number of stations sorted by least
distance: (top to bottom) Merced, CA; San Francisco International Airport, CA; Minden, NE; and
Partoun, UT.
OCTOBER 2005 HUBBARD AND YOU 1529

TABLE 2. Number of stations within a distance.

Merced, San Francisco Minden, Partoun,

Radius CA Intl Airport, CA NE UT
32 km 1 11 4 0
80 km 11 33 18 5
161 km 54 69 68 21

that 10 stations are reasonable for the estimates for

Tmax and Tmin (Figs. 2 and 3). We therefore use 10
stations for the SRT method in the following compari-
sons to the IDW method.
Figures 6 and 7 compare the distribution of rmse for
the SRT and IDW methods for the four states individu-
ally. Figure 6 compares the results for the SRT method
using 10 stations and the best results obtained using the
IDW method (15 stations). The SRT method is supe-
rior to the IDW method for each of the four states for
both Tmax and Tmin. The distribution varies between
states. Nebraska has the least rmse for both the SRT
and IDW methods for Tmax and Tmin. The rmses for the
IDW method for 90% of the stations are 10°, 8°, 6°, and FIG. 7. The proportion of stations (P) with a value of rmse
smaller than the value shown for both the IDW and SRT methods
6°F for Tmax and 8°, 8°, 4°, and 6°F for Tmin for CA, UT, over each of the four states (CA, NE, NV, UT).
NE, and NV, respectively. The rmses for the SRT
method for 90% of the stations are 3°, 2.5°, 2.5°, and
2.7°F for Tmax and 3°, 3.5°, 2.5°, and 3.5°F for Tmin for
CA, UT, NE, and NV, respectively, when 10 best-fit Figure 7 compares the best results obtained using the
stations were utilized to give the estimates. IDW method (15 stations) and the worst results for the
SRT method using only one station. The results of the
SRT method using only one station are still better than
the best results obtained by the IDW method. Thus, the
SRT method outperforms the IDW method.

5. Discussion and conclusions

The parameters for the SRT method were evaluated

in this study. A distance of 80 km for radius of inclusion
is recommended for both Tmax and Tmin. For some iso-
lated stations, the radius of inclusion should be larger
than 80 km, even up to 200 km, so that enough weather
stations are available for the estimation and quality as-
surance processes of Tmax and Tmin. The SRT method
using a time-window length of 60 days gives relatively
good estimates for both Tmax and Tmin, and using this
width is not significantly different from using broader
windows, although broader windows provide somewhat
better estimates. The estimates are better when the
window is centered on the datum of interest. However,
for real-time data quality assurance, we suggest using
FIG. 6. The proportion of stations (P) with rmse less than the
the data time series of previous records or using the
value shown for the SRT and IDW methods to number of stations fixed SRT method.
over each of the four states (CA, NE, NV, UT). The performance of the SRT method stabilized when
 1530 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 22

we used more than 10 stations that were sorted by the ——, P. A. Pasteris, H. F. Diaz, M. S. Plantico, and N. J. Lott,
least rmse between time series of the station of interest 2000: Creating a serially complete, national daily time series
of temperature and precipitation for the western United
and reference stations. There were significant changes
States. J. Appl. Meteor., 39, 1580–1591.
in the performance of the SRT method when fewer Gandin, L. S., 1988: Complex quality control of meteorological
than 10 stations were used. In this study, we find 10 to observations. Mon. Wea. Rev., 116, 1137–1156.
be a minimum but acceptable number of stations for Guttman, N. B., and R. G. Quayle, 1990: A review of cooperative
stabilizing estimates for the SRT method. temperature data validation. J. Atmos. Oceanic Technol., 7,
334–339.
This study shows that the worst estimates using the
——, C. Karl, T. Reek, and V. Shuler, 1988: Measuring the per-
SRT method are superior to the best estimates ob- formance of data validators. Bull. Amer. Meteor. Soc., 69,
tained by the IDW method. Thus, in general, the IDW 1448–1452.
method cannot perform better than the SRT method. Hubbard, K. G., 2001a: Multiple station quality control proce-
The findings provide us more information beyond sim- dures. Automated Weather Stations for Applications in Ag-
riculture and Water Resources Management, K. G. Hubbard
ply confirming that the spatial regression test will be
and M. V. K. Sivakumar, Eds., Tech. Doc. AGM-3 WMO/TD
better than the IDW method (YHG; Eischeid et al. 1074, High Plains Regional Climate Center, Lincoln, NE, 248
2000). We suggest using the SRT method to provide pp.
estimates for any missing data and for QA of the ob- ——, 2001b: The Nebraska and High Plains regional experience
servations. with automated weather stations. Automated Weather Sta-
The unequal distribution of the weather may con- tions for Applications in Agriculture and Water Resources
Management, K. G. Hubbard and M. V. K. Sivakumar, Eds.,
taminate the weighting factor of the spatial regression Tech. Doc. AGM-3 WMO/TD 1074, High Plains Regional
test. For an unequal distribution of input data points, Climate Center, Lincoln, NE, 248 pp.
such as the dense network within the larger-scale net- ——, A. T. De Gaetano, and K. D. Robbins, 2004: A modern
work, the analyzed field may be biased by the data Applied Climatic Information System (ACIS). Bull. Amer.
within the dense network. For example, the coastal sta- Meteor. Soc., 85, 811–812.
——, S. Goddard, W. D. Sorensen, N. Wells, and T. T. Osugi,
tions have only the neighboring stations on the land but
2005: Performance of quality assurance procedures for an
no stations in the sea. Thus, the estimation of data at a Applied Climate Information System. J. Atmos. Oceanic
coastal station may be biased toward the inland sta- Technol., 22, 105–112.
tions, which have significantly different climate pat- Meek, D. W., and J. L. Hatfield, 1994: Data quality checking for
terns than that of the coastal stations. More analysis is single station meteorological databases. Agric. For. Meteor.,
69, 85–109.
planned to address these effects.
Nash, J. E., and J. V. Sutcliffe, 1970: River flow forecasting
through conceptual models. J. Hydrol., 10, 282–290.
REFERENCES
Wade, C. G., 1987: A quality control program for surface meso-
Eischeid, J. K., C. B. Baker, T. Karl, and H. F. Diaz, 1995: The meteorological data. J. Atmos. Oceanic Technol., 4, 435–453.
quality control of long-term climatological data using objec- Willmott, C. J., 1981: On the validation of models. Phys. Geogr., 2,
tive data analysis. J. Appl. Meteor., 34, 2787–2795. 184–194.

View publication stats