See discussions, stats, and author profiles for this publication at: [Link]

net/publication/329961426

An Adjusted Inverse Distance Weighted Spatial Interpolation Method

Conference Paper · January 2018

DOI: 10.2991/cimns-18.2018.29

CITATIONS READS

30 2,436

4 authors, including:

Zhengquan li
China Meteorological Administration
26 PUBLICATIONS 1,031 CITATIONS

SEE PROFILE

All content following this page was uploaded by Zhengquan li on 19 September 2019.

The user has requested enhancement of the downloaded file.

 Advances in Computer Science Research, volume 65
3rd International Conference on Communications, Information Management and Network Security (CIMNS 2018)

An Adjusted Inverse Distance Weighted Spatial

Interpolation Method
Zhengquan Li, Kuo Wang*, Hao Ma and Yaoxiang Wu
Zhejiang Climate Center, Zhejiang Meteorological Bureau, Hangzhou, 310017, China
*
Corresponding author

Abstract—Considering the shortcomings of IDW the method of gradient plus inverse distance squared (GIDS)
interpolation, this study improved the IDW algorithm and was put forward as a modification for IDW method [19-20]. On
proposed a new spatial interpolation method that is called the other hand, an adaptive inverse-distance weight that can
Adjusted Inverse Distance Weighted (AIDW). The AIDW is vary p value of IDW according to sample points clustering
capable of taking into account the comprehensive influence of pattern was also proposed in some studies [14,21]. However,
relative distance and position of sample points on the interpolated the published IDW methods only weigh the influence of
point, by adding a coefficient (K) to IDW formula. The coefficient sample point distance on the estimated point, neglecting the
(K) is used to adjust the distance weight of sample point effect of sample point position on the estimated point.
according to its shielded effect in sample point positions.
Therefore, the interpolating reasonableness of IDW still faces
Theoretical analysis and case study indicates that the AIDW
algorithm could diminish the IDW interpolation error of non-
suspicion, especially on the condition that sample points do not
uniform distribution of sample points, consequently the AIDW evenly distribute surrounding the estimated point.
interpolating is more reasonable, compared with the IDW Considering the comprehensive influence of distance and
interpolating. On the other hand, the contour plotting of the position of sample point on interpolated point, this study brings
AIDW interpolation can effectively avoid the implausible isolated out an improved method of IDW interpolation that is named as
circles and concentric circles that originated from the defect of Adjusted Inverse Distance Weighting (AIDW). In this paper,
the IDW interpolation. The contour derived from the AIDW
firstly introduces the application limit of IDW interpolating
interpolated surface is more similar to the professional manual
method. Next, the assumption and expressions of AIDW is
identification than that from the IDW interpolated surface.
given and the advantage of AIDW is described by theoretical
Keywords—AIDW; spatial interpolation; contour plotting; IDW analysis. And then, the AIDW performance is analyzed by
conducted a case study. Lastly, the study gives a brief
I. INTRODUCTION conclusion of AIDW application.
A basic issue in spatial analysis is interpolating a II. ADJUSTED INVERSE DISTANCE WEIGHTED
continuous surface from point samples. Numerous spatial
interpolation methods have already been developed for A. Shortcoming of inverse distance weighted
supporting transformation from point data to continuous IDW interpolation accords to the First Law of Geography
surface map [1-5]. Those methods can be divided into "everything is related to everything else, but near things are
geographical statistics, non-geographical statistics and hybrid more related than distant things." [13,22]. The IDW formula is
approach. Natural neighbor, splines and trend surface given as (1).
polynomial regression, radial basic function (RBF), triangular
irregular network (TIN) and inverse distance weighted (IDW), n n
1 1
which are all classified as non-geographical statistics [1,6]. Z  Zi  (d ) (1)
Various forms of Kriging methods can be classified as i 1 (di ) p i 1 i
p

geographical statistics [7-8]. Hybrid interpolation approach

includes regression Kriging, linear mixed model, lapse rate
Where Z is the estimated value for prediction point, Zi is
combined with Kriging, and so on [9-12].
the measured value for sample point, di is the Euclidean
Probably one of the most famous interpolation methods is distance between sample point and prediction point, p is a
IDW, which is a simple and intuitive deterministic power parameter, and n represents the number of sample points.
interpolation. IDW is easy to be implemented and available in A main factor affecting IDW interpolation result is the p value.
almost any GIS software, so it is applied frequently in various When the p value increases, the smoothness of IDW output
disciplines [1]. The well-known advantage of IDW is that surface increases [23].
keeping the measured value at sample location [5,13]. But IDW
It is well known that the performance of IDW interpolation
has a shortcoming that it always does not work well on
is better when sample points distribute regularly, but it faces
conditions when the sample points are clustering [14-15].
various suspicions if sample points are clustering, because high
Moreover, many studies have reported that IDW interpolation
clustering of sample points will reduce the accuracy of IDW
is inclined to emerge "bull's eye" patterns in contour plotting
interpolation [24-26]. Moreover, IDW interpolating surface is
[16-18]. Some suggestions for enhancing IDW method have
sensitive to the presence of outliers, and it suffers from
been proposed as well. For improving interpolation accuracy,

Copyright © 2018, the Authors. Published by Atlantis Press.

This is an open access article under the CC BY-NC license ([Link] 128
 Advances in Computer Science Research, volume 65

discontinuities at sample points resulting in peaks or troughs,

especially where sample points are sparse. As a result, the
contour map derived from IDW interpolated surface usually
shows many isolated circles and concentric circles (commonly
called as "bull's eye"), which affects the aesthetic property of
contour map [17-18]. To avoid the shortcomings of IDW
interpolation, this paper improves IDW algorithm and puts
forward an AIDW interpolation method.
B. Assumption of AIDW FIGURE I. EXAMPLES OF SAMPLE POINTS SCATTERING IN
INTERPOLATING NEIGHBORHOOD (FIVE POINTS)
AIDW is similar to IDW abiding by the First Law of
Geography. A key difference between AIDW and IDW is where sinpθij is used to weigh the shielded influence of i
adding a coefficient (K) in IDW formula to adjust the distance sample point by j sample point (1≤j≤i-1). In (3), ki is expressed
weight of sample point. The coefficient (K) is based on an as a multiplicative model of sinpθij. As a result, the more
assumption that sample point which is close to prediction sample points shield i point, the lower ki value is, and the
location has shielding influence on those sample points which interpolating weight of i point decreases much more. Due to the
are farther away to prediction location, and the shielding closest point to prediction location is not shielded by any other
influence can reduce the distance weight of shielded sample sample point, so k1 is set to 1.
point. The shielding influence appears when α<360º/n, and
disappears when α≥360º/n. α is an intersection angle for two The sinpθij can be calculated by (4).
lines which connect prediction point and two sample points
(Figure I), and n is the number of sample points used for the
 1  ij  360 / n
prediction. Figure Ia shows that all sample points distribute 
evenly surrounding prediction point (black). The intersection sin ij  
p
p (4)
angle (α) of random two sample points is greater than or equal (1  cos 2 ij ) 2  ij  360 / n
to 72 degree (α≥360º/5). AIDW assumes that these sample
points have no shielding or shielded influence. In Figure Ic,
sample points are high clustering and the maximum where θij is the intersection angle formed by the line that
intersection angle (α) is lower than 72 degree. On this condition, connects i and j sample points and the line that passes through
AIDW assumes that Z1 have shielding influence on Z2, Z3, Z4 prediction point and the midpoint of line 1. αij is the
and Z5, Z2 influences Z3, Z4 and Z5, Z3 influences Z4 and Z5, intersection angle for two lines that connect prediction point, i
and Z4 influences Z5. In most cases, sample points distributing and j sample points separately. θij and αij are illustrated in
is more analogous to Figure Ib, neither like Figure Ia nor Figure II.
Figure Ic. The intersection angle (α) of Z1 and Z3 is equal to 72 Integrating (2) with (3), AIDW algorithm can also be
degree in Figure Ib, so AIDW assumes that Z1 and Z3 have no expressed as the following.
shielding influence on other sample points, but Z2 influences Z4
and Z5, moreover Z4 influences Z5. Z1 Z 2 Z Z
 sin p  21  3p sin p 31 sin p 32    np sin p  n1 sin p  n 2 sin p  n3  sin p  n ,n 1
d 1p d 2p d3 dn
C. Expressions of AIDW algorithm Z
1 1 1 1
 sin p  21  p sin p  31 sin p 32    p sin p  n1 sin p  n 2 sin p  n3  sin p  n ,n 1
Combining the IDW formula, AIDW algorithm can be d 1p d 2p d3 dn
expressed as (2). (5)
n n
ki ki
Z  Zi  (d ) (2) D. Capability of weighing shielded influence
i 1 (di ) p i 1 i
p
In course of Zi moving to prediction location (Zo) with a
constant angle (Figure IIa), θ and sinpθ values are increasing
Where Z, Zi, di and n is the same as that of the (1), i is a which indicates that the shielding influence of Zj on Zi is
sequence number of sample point which is sorted by the di gradually reducing. When Zi moves to the point whose distance
value (from low to high), and ki is an adjusted coefficient of between Zi and Zo is equal to the distance between Zj and Zo,
distance weight which represents the comprehensive shielded the three points (Zo, Zj and Zi) form an isosceles triangle. On
influence of i sample point. The ki is calculated by (3), whose this condition, θ is 90 degree, and sinpθ value is l, which means
value ranges from zero to 1. that Zj has no shielding influence on Zi anymore.
In process of Zi moving to right behind Zj with a constant
 1 i 1 distance between Zi and Zo (Figure IIb), θ and sinpθ values are
 j i 1
ki   (3) decreasing, which indicates the shielding influence of Zj on Zi
  sin ij i  2,3, 4  n
p
is gradually enhancing. If θ and sinp θ value is 0 which means
 j 1 Zj completely shields Zi. That is to say, Zi is useless for
estimating Zo.

129
 Advances in Computer Science Research, volume 65

Regarding the moving places of Zi as the locations of other In order to concisely compare contour plotting of IDW and
sample points, Figure II illustrates the variation of θ and sinpθ AIDW interpolation, a simply example is given in Figure VI.
in two extreme cases of sample point distribution. One case is The distance of purple point to prediction point (black point) is
that the distances between two sample points and prediction same and its value is l. Tawny points lie right behind purple
point are tending to be equal (Figure IIa), θ → 90º， sinpθ→ 1. points, and their distances to prediction point are also same
The other case is that one sample point is tending to right (whose value is 2). The value of prediction point estimated by
behind the other points (Figure IIb), θ → 0º, sinpθ→ 0. IDW (setting p=1) is 50, while the value estimated by AIDW
(setting p=1) is 20, because tawny points are shielded
completely by purple points which causes the weights of tawny
points reducing to zero. From this simply example, it can be
seen clearly that the interpolating result of AIDW is more
reasonable than IDW.

FIGURE II. EXAMPLES OF sinpθ VALUE CHANGING WITH SAMPLE

POINT MOVING

The two extreme cases can cover all position relationship of

random two sample points; therefore it is reasonable using FIGURE III. DIFFERENT TREATING MODE OF SAMPLE POINTS IN
sinpθ to weigh the shielding or shielded influence of sample IDW AND AIDW INTERPOLATION METHODS
points.
E. Characteristic of AIDW interpolation
By adding the adjusted coefficient (K) in IDW formula,
AIDW can consider the comprehensive influence of distance
and position of sample point. When the distribution of sample
points is even (Figure Ia), or the distances of all points to
prediction point are same, the interpolating result of AIDW is
the same as that of IDW interpolating, so IDW can be regarded
as a special case of AIDW. When sample points are clustering
in a section such as Figure III, AIDW can make sample points
that are used for interpolating tend to even distributing. On the
FIGURE IV. CONTOUR PLOTTING EXAMPLE OF IDW AND AIDW
other hand, AIDW is capable of reducing the probability of INTERPOLATION METHODS (THE VALUE OF RED AND GREEN
"bull's eye" emerging in contour plotting (Figure VI). CONTOURS IS 30 AND 40, RESPECTIVELY)
Figure IIIa shows sample points which are high clustering
in right-upper section of prediction point (black point), the As shown in Figure VIa, since the low values of purple
result of IDW interpolating cannot be guaranteed due to points surround the higher value of prediction point estimated
contravening the principle of geo-statistical analysis that by the IDW, the chance of the concentric circles pattern
sample points are even distribute. However, considering the emerging would increase in contour plotting. However, the
shielding or shielded influence of sample points, AIDW value of prediction point estimated by the AIDW is close to the
interpolating will reduce the distance weights of Z6, Z7 and Z8 values of purple points, therefore prediction point and purple
(Z6 shielded by Z1, Z7 shielded by Z3 and Z4, and Z8 shielded points could be classified into a same zone, avoiding the
by Z3), and make the distance weights of Z10, Z11 and Z12 to be appearance of the concentric circles pattern (Figure VI b). As
zero (Z10, Z11 and Z12 is completely shielded by Z4, Z7 and Z8, for a case study of AIDW and IDW in contour plotting, it is
respectively), as shown in Figure IIIb. Consequently, the value given in Section III.
of prediction location is estimated by (6). Because the AIDW III. CASE STUDY OF AIDW AND IDW INTERPOLATION
interpolating of prediction point mainly considers purple points
and partly considers green points (Figure IIIb), AIDW has Based on annual precipitation data from 71 weather stations
made sample points tend to distribute evenly. in Zhejiang province, this study had interpolated precipitation
surface by individually applying IDW and AIDW method. In
the interpolating process, setting the power parameter (p) was 2,
Z1 Z 2 Z 3 Z 4 Z 5 Z 6 sin p  61 Z 7 sin p  73 sin p  74 Z8 sin p 83 Z 9
        p interpolating neighborhood included 5 weather stations, and the
d1p d 2p d3p d 4p d5p d 6p d 7p d8p d9
Z spatial resolution of precipitation surface was l km × l km.
1 1 1 1 1 sin  61 sin  73 sin  74 sin 83 1
p p p p
        p IDW interpolating and its contour plotting were conducted by
d1p d 2p d3p d 4p d5p d 6p d 7p d8p d9
(6) ArcMap9.3 software, and AIDW was conducted by the
programs mentioned above. The errors of the IDW and AIDW
interpolating for annual precipitation are listed in Table I.

130
 Advances in Computer Science Research, volume 65

Compared to IDW interpolation, the result of AIDW weight of KH station decreases with prediction location being
interpolation is more reliable. The errors of AIDW far away from KH station, meanwhile the sum weight of its
interpolation are all lower than that of IDW interpolation, surrounding station is relatively increasing. As a result, the
especially in the mean error and the root mean square error. further away from KH station, the smaller an estimated value
of prediction location is, which results in a "bull's eye" pattern
For the contour plotting, though the surface and contours surrounding KH station (Figure Va). In the interpolating
resulted from AIDW method have a broadly similar pattern process of AIDW method (especially for the interpolations of
with that from IDW method, there are significant differences in the left area of KH station), KH station has shielding influences
local areas marked by the rectangular (Figure V). The "bull's on CS, JS, QZ and CA station, and the four stations have also
eye" (in the red rectangular) and isolated circle (in the black shielding or shielded influences on each other. Consequently,
rectangular) in Figure V a are disappeared in Figure Vb. the weights of surrounding stations are reduced relatively and
the estimating of prediction location is mainly controlled by
TABLE I ERRORS ANALYSIS OF CROSS VALIDATION OF THE IDW
AND AIDW INTERPOLATION
KH station. Since little difference among the estimated values
in the left area of KH station, these estimated values are
Mean Mean Mean Root Mean classified into a same zone which avoids "bull's eye" emerging
Interpolation Error Absolute Relative Square in contour plotting (Figure Vb). In those areas marked using
Method /mm Error Error Error
/mm /% /mm letter B, C or D, the interpretation about the differences
IDW 8.2 95.9 7.1% 179.8 between IDW and AIDW interpolating is analogous to that in
AIDW 3.9 87.3 6.5% 136.3 the area A.
M
IDW C Besides reducing the probability of "bull's eye" emerging,
B AIDW has a capability to decrease the chance of isolated circle
H
appearing. In the area marked by letter M, there is an isolated
circle surrounding CX station (Figure Va). Owning to HZ and
DQ stations existing, the value estimated by IDW method
N
becomes greater while the prediction location is further away
from CX station, which results in the appearance of isolated
A circle. In the interpolating process of AIDW method (especially
Legend
Unit: mm
for the interpolations of the left area of CX station), HZ station
K
S is subjected to the shielding influence of CX station, and DQ
station suffers the corporative shielding influences of CX, HZ
D
and AJ station. Therefore, the contribution of DQ station to the
estimated value is greatly depressed, and HZ station is partly
T
decreased too. As a result, the interpretation of the left area of
CX station is mainly dominated by CX and AJ stations, which
results in the similar interpolated values, so the isolated circle
a does not appear in Figure Vb. In those areas marked using
letter N, S or T, the interpretation for appearing or disappearing
M
AIDW C of the isolated circle is analogous to that of the area M.
B H It is worth noting that AIDW has not eliminated the "bull
eye" at DQ station and isolated circle at LY station which
marked by the yellow rectangular frame (Figure V), because
the four stations surrounding DQ or LY station distribute
N
evenly (one in each quadrant area). It demonstrated that AIDW
A is similar to IDW when the distribution of sample points is
Legend
Unit: mm even.
K
S
Generally speaking, the precipitation surface interpolated
by AIDW method is more reasonable than IDW method,
especially for the interpolations of the surface frontier areas.
D
Moreover, AIDW method can reduce the number of
T
implausible isolated circle and concentric circles objectively,
and its contour plotting is smoother comparing with IDW
method.
IV. CONCLUSION AND DISCUSSION
FIGURE V. COMPARISON OF CONTOUR PLOTTING OF IDW AND
AIDW INTERPOLATION METHODS Though IDW method is commonly used in spatial analysis,
it has the shortcomings of neglecting sample points clustering,
In the area marked by letter A, the annual precipitation of and inclining to emerge isolated circle and concentric circles in
KH station is much higher than that of its surrounding stations its contour plotting. For the sake of overcoming above
(CS, JS, QZ and CA station). According to IDW algorithm, the problems, this study had improved IDW algorithm and put

131
 Advances in Computer Science Research, volume 65

forward a new method called AIDW. According to the [8] H. C. Yeh, Y. C. Chen, C. Wei, and R. H. Chen. “Entropy and kriging
intersection angle (α), AIDW decides whether or not to have approach to rainfall network design,” Paddy Water Environ, 2011, 93,
343-355.
shielding influences among sample points, and weighs the
[9] X. Li, G. Cheng, and L. Lu. “Spatial analysis of air temperature in the
shielded contribution of the point shielded by other points Qinghai-Tibet Plateau,” Arct Antarct Alp Res, 2005, 37, 246-252.
using the coefficient (K). Theoretical analysis and the case [10] W. Z. Shi and Y. Tian. “A hybrid interpolation method for the
study indicated that AIDW method could acquire more refinement of a regular grid digital elevation model,” Int J Geogr Inf Sci,
reasonable interpolations than IDW method. Moreover, the 2006, 20, 53-67.
contour plotting of AIDW can effectively avoid the implausible [11] E. Ozelkan, S. Bagis, E. C. Ozelkan, et al. “Spatial interpolation of
isolated circles and concentric circles which often exist in IDW climatic variables using land surface temperature and modified inverse
interpolating; its contour plotting has more rationality and distance weighting,” Int J Remote Sens, 2015, 36(4), 1000-1025.
aesthetic. [12] Y. H. Wang, H. B. Yang, D. W. Yang, et al. “Spatial interpolation of
daily precipitation in a high mountainous watershed based on gauge
Due to involving the calculation of α and sin θ, the p observations and a regional climate model Simulation,” J Hydrometeorol,
computing efficiency of AIDW method is slight lower than that 2016, 18(3), 845-862.
of IDW method. In the above case study, the computational [13] K. Johnston, J. M. V. Hoef, K. Krivoruchko, and N. Lucas. “Using
ArcGIS geostatistical analyst,” New York: Environmental systems
time of either AIDW or IDW was less than l s. But the running research institute, Inc. 2001.
time of AIDW method is increasing more than IDW method,
[14] G. Y. Lu and D. W. Wong. “An adaptive inverse-distance weighting
while the interpolating resolution is higher and interpolating spatial interpolation technique,” Comput Geosci, 2008, 34, 1044-1055.
neighborhood including more sample points. When [15] M. Azpurua and K. D. Ramos. “Comparison of spatial interpolation
interpolating resolution was 500 m × 500 m and the methods for estimation of average electromagnetic field magnitude,”
interpolating neighborhood included 15 weather stations, Prog Electromagn Res M, 2010, 14, 135-145.
AIDW method had cost 11 s to conduct interpolating [16] C. A. Gotway, R. B. Ferguson, G. W. Hergert, and T. A. Peterson.
precipitation surface of Zhejiang province, and IDW method “Comparison of kriging and inverse distance methods for mapping soil
consumed 7 s to do the same thing. However, it is worth getting parameters,” Soil Sci Soc Am J, 1996, 60, 1237-1247.
the interpolating reasonability regardless of the slight more [17] Z. M. Feng, Y. Z. Yang, X. Q. Ding, and Z. H. Lin. “Optimization of the
spatial interpolation methods for climate resources,” Geogr Res, 2004,
running time. 23, 357-364.
ACKNOWLEDGMENT [18] D. Nusert and S. Dug. “Applying the inverse distance weighting and
kriging methods of the spatial interpolation on the mapping the annual
This study was supported by the Public Welfare precipitation in Bosnis and Rzegovina,” In: R. Seppelt, A. A. Voinov, S.
Technology Research Projects of Zhejiang Province Lange, D. Bankamp. International Congress on Environmental Modeling
and Software. Leipzig: iEMSs. 2012.
(2015C33055), Natural Science Foundation of China
(41605049), and Fund for Meteorological Science and [19] J. Hooyberghs, C. Mensink, G. Dumont, and F. Fierens. “Spatial
interpolation of ambient ozone concentrations from sparse monitoring
Technology of Zhejiang Province, China (2017ZD13, points in Belgium,” J Environ Monitor, 2006, 8, 1129-1135.
2017QN04). [20] N. S. Golkhatmi, S. H. Sanaeinejad, B. Ghahraman, et al. “Extended
modified inverse distance method for interpolation rainfall,” Inter J Eng
REFERENCES Inv, 2012, 1, 57-65.
[1] J. Li. and A. D. Heap. “A review of spatial interpolation methods for [21] G. Y. Lu, L. S. Chiu, and D. W. Wong. “Vulnerability assessment of
environmental scientists,” Canberra: Geoscience Australia, 2008. rainfall-induced debris flows in Taiwan,” Nat Hazards, 2007, 43, 223-
[2] M. Delbari, P. Afrasiab, S. Jahani. “Spatial interpolation of monthly and 244.
annual rainfall in northeast of Iran,” Meteorol Atmos Phys, 2013, 122(1- [22] W. R. Tobler. “A computer movie simulating urban growth in the
2), 103-113. Detroit region,” Econ Geogr, 1970, 46, 234-240.
[3] S. Wang, G. H. Huang, Q. G. Lin, et al. “Comparison of interpolation [23] R. B. Husar, and S. R. Falke. “Uncertainty in the spatial interpolation of
methods for estimating spatial distribution of precipitation in Ontario, PMlO monitoring data in Southern California,” [online], 1996,
Canada,” Int J Climatol, 2014, 34(14), 3745-3751. [Link]
[4] P. D. A. Borges, J. Franke, H. Weiss, and C. Bernhofer. “Comparison of [24] D. Zimmerman, C. Pavlik, A. Ruggles, and M. P. Armstrong. “An
spatial interpolation methods for the estimation of precipitation experimental comparison of ordinary and universal kriging and inverse
distribution in distrito federal, Brazil,” Theor Appl Climatol, 2016, distance weighting,” Math Geol, 1999, 31, 375-390.
123(1-2), 335-348. [25] K. Stahl, R. D. Moore, J. A. Floyer, M. G. Asplin, and I. G. Mckendry.
[5] T. Chen, L. L. Ren, and F. Yuan. “Comparison of spatial interpolation “Comparison of approaches for spatial interpolation of daily air
schemes for rainfall data and application in hydrological modeling,” temperature in a large region with complex topography and highly
Water, 2017, 9(5), 342-342. variable station density,” Agr Forest Meteorol, 2006, 139, 224 - 236.
[6] F. W. Chen and C. W. Liu. “Estimation of the spatial rainfall distribution [26] K. Wang, H. Ma, J. Li, B. H. Gu, and H. Wu. “Assessment of the
using inverse distance weighting (IDW) in the middle of Taiwan,” Paddy POEM2 model for simulating tropical intraseasonal oscillation,” J Trop
Water Environ, 2012, 10, 209-222. Meteorol, 2018, 24(3): 323-333.
[7] H. Apaydin, F. K. Sonmez, and Y. E. Yildirim. “Spatial interpolation
techniques for climate data in the gap region in Turkey,” Clim Res, 2004,
28, 31-40.

132
View publication stats