ISSN 2321–8355 IJARSGG (2014) Vol.2, No.

1, 45-61
Research Article

International Journal of Advancement in Remote Sensing, GIS and

Geography

GEOSTATISTICAL APPROACH IN SPATIAL

ESTIMATES OF RAINFALL: BETWA RIVER
CATCHMENT
I
Ritu Ahlawat
I
Department of Geography, Miranda House, University of Delhi, INDIA ([email protected])
----------------------------------------------------------------------------------------------------------------------------- -----
ABSTRACT: As the point-based rainfall observation design has implications in correctly representing spatially
continuous rainfall values, therefore methods ranging from simple arithmetic mean to isopleths and Thiessen’s polygon-
based weighted means have been generally employed by hydrologists in order to obtain mean basin rainfall. The utility of
these methods is important especially for analysing the spatial distribution of network and thereafter to suggest locations.
Hence, advanced mathematical versions using network theory, geostatistical approach and optimum path analysis have
been used to select the optimum design even in the preliminary stage of investigation. At a catchment level, an application
of various interpolation methods using geostatistics is demonstrated here in case of Betwa river, one of the important
southern tributary of Yamuna river after Chambal. The script for these methods was written in Geographical Information
System (GIS) Software – ILWIS (Integrated Land and Water Information System) and it was run for normal rainfall
parameters using data from Indian Meteorological Department. In the geostatistical analysis, it was found that trend
surface, moving average and kriging based interpolated map provided good estimates of spatial accuracy of mean basin
rainfall within 10-12 % of coefficient of variation when compared with normal rainfall at control stations. Among kriging
methods, ordinary kriging was found to be best in terms of least root mean square deviation from validation results. Error
map based on anisotropic kriging resulted into least amount of absolute error within 10% of estimation in rainfall from
the measured normal values. The error can be further reduced if more rain-gauges are located in spatial dispersion zones
by taking into account the effect of direction and elevation on rainfall.

KEYWORDS: Spatial interpolation, reflexive nearest neighbour, semi-variogram, anisotropic trend, error rainfall maps
----------------------------------------------------------------------------------------------------------------------------------
1. INTRODUCTION: Scale is a key issue in hydrology and the catchment is considered to be the basic unit of
measurement. Large-scale design of any monitoring network or planning for water resource project has to overcome
some of the point-based sampling difficulties. The decision on operational number of sites of observation is based on
the level of accuracy required and the spatial variability of the area or variable under investigation. Till the mid 60’s
and 70’s, the design of hydrological data collection network has been mostly based on minimising cost, which
translates into problem of accessibility and maintenance of observation status. But, the present day hydrological
capabilities demand high level of accuracy in the statistics obtained from collected data. Not only long-term record,
but also spatial representatives of the data is essential to provide reliable inputs for various purposes like-agricultural
planning, crop sequence and the acreage, irrigation schedule and canal network, capacity of reservoirs, sustainability
of demand for groundwater resources, means for short-term water supply, quality of water used for various purposes,
etc. Microclimate, local relief, changing land use pattern may result in significant spatial variations of a point-based
estimate of the hydrological phenomena (1).
The issue of spatial variation in estimates of mean basin rainfall has been well recognized now and is being
addressed scientifically. Besides, simple arithmetic mean to isopleths and Thiessen’s polygon-based weighted means,
the more mathematical ones using determinants and matrices have been used to determine area of small triangles
around a rain gauge. However, due to subjectivity involved in selection of nearest station to create Theissen polygon
or triangular mesh, the trial and error method does not differ significantly from the simple arithmetic method so far as
mean estimates are concerned. The need to recognise accuracy in the spatial representation of hydrological data has
given rise to more sophisticated statistical analysis. Configurations of spatial network based on an estimation of time-
averaged areal mean of precipitation have been analysed particularly with respect to their location, duration of
observation and measurement of error (2).Use of geostatistics, initially in mining field, started in hydrological
investigations too. These spatial data based computations based on kriging methods have been employed to reduce the
error of representativeness in estimation of mean basin parameters.
Kriging, named after Krig (3), can be seen as a method of point interpolation. The weight factors in kriging
are determined by using a user-specified semi-variogram model (based on the output of the spatial correlation
operation), the distribution of input points, that are calculated in such a way that they minimize the estimation error in
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 46

each output pixel. Semi-variogram is a basic geostatistical measure to determine the rate of change of a regionalized
variable along a specific orientation (usually distances). First, the distances between all points pairs which fall within
the same user-specified lag, i.e. the same distance (and direction) class are calculated, i.e. omni-directional and then
bi-directional spatial correlation is calculated to ascertain the directional trend (anisotropy) of rainfall by performing a
variogram surface operation. The computation of variograms of positional uncertainty is possible when measured by
geometrical discrepancy. These variograms prove to be efficient tool to represent spatial autocorrelation (4). In an
another study incorporatting the effect of elevation into spatial interpolation of rainfall in a watershed of Portgual
showed that larger prediction errors were obtained based on both Thiessen polygon and inverse square distance
method. Ordinary kriging yields more accurate predictions than linear regression when the correlation between rainfall
and elevation is moderate (5).
In India, however, the use of geostatistical approach to arrive at mean estimates of basin rainfall is limited.
Here, mean basin rainfall estimates based on Thiessen polygon and isohyetal method are quite common using the
existing network of rain-gauging stations. A locational analysis of the network based on annual data in the case of
Dhansiri basin of north-east India stressed upon the need of accuracy level in regional pattern of rainfall. Based on
coefficient of variation of rainfall, an optimum network was determined for the basin at 5% error in estimation. But,
exact location of the new rain gauge was not proposed (6). Location of proposed rain-gauging network, considering
the orientation of sub-catchments, distances among existing rain-gauge sites, accessibility and proximity to nearby
river-gauge and discharge sites, has been attempted in another river basin in the east, the Mahanadi (7). However, the
method involving the correlation coefficients to compute optimum distances was not illustrated in detail.
Thus, several approaches have been used for estimating regional values from point measurements ranging
from simple arithmetic average method to kriging based surface fitting methods. In the light of paucity of works
related to correct estimation of mean basin rainfall, spatial interpolation methods need to be intensively explored for
better results. The present study seeks to analyse variability in spatial pattern of rainfall at catchment level by
interpolation of rainfall data using various geostatistical methods. A comparison of these methods can help decision
makers of water resources to choose suitable method of interpolation and in understanding spatial implications of
variability in mean basin rainfall.

2. RESEARCH AREA: Betwa catchment, lying between the 77 o10 East to 80o20 East Longitudes and 22o50 North
to 26o00 North latitudes, is selected for the present study. It is of rhomboidal shape and covers an area of about 43
thousands km2, bounded by Vidhyanchal Scraplands on the south, Malwa plateau on the southwest, Chambal-Sindh
ravines on the west and the catchment area of Ken river basin marks its eastern boundary (Fig. I a). The Betwa river
rises in the Raisen district of Madhya Pradesh near village Barkhari, southwest of Bhopal at an elevation of about
576m above mean sea level and joins Yamuna river at Hamirpur in southern Uttar Pradesh at an elevation of about
101 m above mean sea level. While flowing in a generally northeastern direction, it cuts through lava traps and
exposes upper Vindhyan rocks. In Budelkhand region near Jhansi it cuts across granite and gniesses giving way to
large reservoirs like Matatila and thereafter it flows through alluvium tract as seventh order river (Fig. I b). Betwa
being an inter-state river follows border of Uttar Pradesh and Madhya Pradesh to some extent and, in all, it covers
(including parts) four districts of southern Uttar Pradesh (Jalaun, Hamirpur, Jhansi and Lalitpur) and ten districts of
Madhya Pradesh (Datia, Shivpuri, Guna, Tikamgarh, Chhatarpur, Sagar, Vidisha,. Bhopal, Sehore and Raisen). The
important sub tributaries are: Bina, Narain, Jamni, Dhasan and Birma on right bank; Kalisote, Halali, Bah, Sagar,
Naren and Kethan on left bank. The catchment area experiences overall higher mean annual temperature ranging from
24 oC to 26.5 oC. Extremes of temperatures show that temperatures can go as high as up to 48 oC and as low as up to
2oC. The Banda district often registers the largest number of sunstroke every year, probably, owing to tense terrestrial
radiation and lack of haziness in the sky being away from the source regions of dust storms in the west (8). The
plateau region experiences a slightly lower range of maximum temperature than that of plains but minimum
temperature follows the reverse pattern. Climographs for the region move towards scorching and muggy conditions in
the pre-monsoon and monsoon season, respectively (9). Normal annual rainfall varies from 788 to 1358 mm in the
catchment area having a network of 58 rain-gauge stations located inside the catchment area and 12 nearby gauges
according to state climatological records of Uttar Pradesh and Madhya Pradesh. Mean rainfall is 981 mm with annual
variability of 16.8%. About 75 cm (or 75%) of precipitation is concentrated in three months only but the water budget
is highly variable with deficit (800 to 1200 mm) surpassing the surplus depth of water (0 to 325 mm) (10). Medium
black soil in upper Vindhyan reaches, mixed red and black soil in middle reaches, alluvial soil in lower plains
characterize the region with dry deciduous type of vegetation. 'Tendu' trees are found in abundance in the Panna and
Chhatarpur districts. Only Panna and Raisen districts have more than 1/3rd of their total reported area under forest
cover. There is very little forest cover (<10%) in the northern plains where the original cover has almost been removed
to make room for cultivation.
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 47

Fig. I a

Fig. I b

Fig. 1 (a & b): Location of Research Area (Betwa Catchment)

(Source: Author using ILWIS based on SOI Toposheets and NATMO Atlas)

3. METHODOLOGY: Long-term annual data for meteorological observatories and rain gauges stations was obtained
from Indian Meteorological Department (IMD). Besides annual data, monthly and weekly data was also obtained from
state-level climatological reports. Further, primary data regarding the nature and functioning of data stations,
maintenance, communication and publication of data, and the economic or other managerial problems, was also
gathered at some of the selected sites.
The boundaries of the Betwa catchment were drawn according to relief and drainage characteristics as
depicted on Survey of India toposheets and Watershed Atlas. Then, location data of rain-gauges was placed on base
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 48

map. According to latest published records of meteorological observatories in Indian Meteorological Department
Report (1999) on Weekly Probabilities of Rainfall, stations having more than 30 years of record (1951-80) were
included to get estimates of mean values at 45 rain gauges out of total 58 located inside Betwa river catchment. From
these 13 stations, mostly having short-term record, 11 were used for validation of results. First, computations of their
areal means based on simple arithmetic average and Theissen polygons were done. Thereafter, spatial distribution of
rain-gage stations was analysed based on point-pattern statistics which was compared to the situation of complete
spatial randomness (CSR).Various isohyetal maps were drawn in a GIS system by writing a script in ILWIS
(Integrated Land and Water Information System) for annual values. Both deterministic and stochastic interpolations
methods were used (Fig. II) to get pixel-wise values by fitting different surfaces. Coefficient of variability (Cv) of
rainfall and standard error was calculated to compare best estimates of mean rainfall data. Further, a detailed analysis
of kriging based operation was performed based on spatial auto-correlation and semi-variogram operations. -
Geostatistics software was also used to get parameter values for semi-variograms.

Methodology

Catchment Base Map Input Data

(SOI Toposheets) (Normal Annual Rainfall)

Location of Rain Gauge Point Attribute Map

Station (Point Map)
Point Interpolation Methods
Direct Weighted
Average Point Pattern
Analysis

Arithmetic
Average Deterministic Stochastic

Nearest Point Moving Average Semi- Spatial

(Theissen Polygon) Variogram Autocorrelation
(Linear Decerease &
Inverse Distance)
Kriging
Trend Surface
(Simple, Ordinary,
(Plane & Polynomial) & Universal)

Moving Surface Anisotropic

(Plane & parabolic) Kriging

Fig. II: Flow Chart of Methodology

By plotting the answers on autocorrelation against the distance classes, limiting distance was calculated to see
until which distance spatial autocorrelation exists between point pairs. This value was used for point interpolations
such as moving average, moving surface and kriging. Then, for all pairs of points that have a certain distance to each
other, the Moran's I (the product of difference of point pair values to the overall difference) and Geary's C (the squared
differences of point pair values to the mean of all values for these point pairs) values were calculated within each
distance class. Thereafter, for the point pairs in a certain distance class and in the correct direction, experimental semi-
variogram values were calculated. By finding a model (Spherical, exponential, Gaussian and wavy semi-variogram
models) which fits these experimental semi-variogram values, necessary input information such as model type, sill,
range, and nugget) was obtained for kriging operations. For Anisotropic kriging, semi-variogram values for two
perpendicular directions were calculated, then zonal anisotropy angle and ratio of the two bands was used to get
anisotropic kriging map. Finally, for all kriging methods error maps were obtained in order to choose best
interpolation method.
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 49

4. RESULTS AND DISCUSSIONS: As estimation of basin rainfall depends on network of rain gauges, therefore
status of existing ones was assessed first. Although at every tehsil headquarter has rain-gauge located within its
premises but during field survey it was observed that it remains operational during monsoon season only due to
official state policy of maintaining rainfall records. Agricultural centres also have their own network of agri-met
observatories. Most of the rain gauges, manned by part-time staff are maintained by Meteorological office from
where a regional advisory is given. Automatic rain - gauges are located at meteorological stations only and even they
are sometimes not fully functions. The overall network of 58 rain gauges is way behind minimum density of 575
Km2 per gauge as per recommendations of World Meteorological organization (WMO) required in the interior
plains and undulating or hilly regions like Betwa Catchment (Fig. III). Here, too, the spatial distribution of
existing network of rain gauges is to be analysed first in order to have justification for locations of proposed
gauges.

Fig. III: Rain-Gauge Stations in Study Area and their Influence Area Theissen Polygons
4.1. SPATIAL DISTRIBUTION OF RAIN GAUGE NETWORK: Spatial distribution of rain gauges with respect
to their first nearest neighbour in the Betwa river catchment exhibits regular pattern than that of present in a situation
of complete spatial randomness pattern - both in terms of number and distance. At some of the higher orders, however,
the neighbourhood statistics shows a little clustered pattern at 2 nd and 5th order, and a regular pattern regarding
number of gauges but a clustered pattern is observed with respect to their distance at all levels (Table I). Thus,
meteorological-cum-rain gauge stations at higher hierarchical levels can be located randomly in space. It is only at
lower levels, the spatial correlation effects need to be examined in detail.

Order Number of Nearest Neighbours Distance to Nearest Neighbours

Observed Assumed Observed Assumed
Values with CSR Values with CSR
1 28 29.21 22757.85 22703.46
2 18 15.47 30303.24 34055.19
3 8 11.43 36908.84 42568.99
4 6 9.47 42239.14 49661.55
5 10 8.27 47478.47 55873.22
6 6 7.44 52704.97 61458.27
Table I: Reflexive Nearest Neighbour Statistics of Rain Gauge Distribution in Betwa River Catchment
(Source: Computed in ILWIS from location data of rain-gauges)
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 50

4.2. BEST ESTIMATES OF MEAN BASIN RAINFALL: Beginning from simplest measure of mean basin rainfall, a
detailed geostatistical analysis using GIS software revealed pixel-wise information of rainfall that can be used for any
study like digital terrain model correlation with rainfall estimates, water balance etc. based on interpolated values. An
overview of different methods along with their results is presented in the following sections.

4.2.1. DIRECT WEIGHTED AVERAGES: The arithmetic average puts the mean annual rainfall in the Betwa
catchment as 969.53 mm. But, this approach provides reasonable estimates if, the gauges are distributed uniformly and
the topography is flat which, except for the northern region, are not found in abundance in the Betwa catchment.
Hence, mean precipitation as 1033.22 mm computed by Thiessen polygon method (Fig. III and Table II) works out to
be higher than that of by arithmetic average as it takes into account local variations.
Out of 58 polygons drawn in Figure III, leaving aside influence area of 12 rain gauges located outside Betwa
catchment, only 9 have area comparable with the optimally sized hexagons having a uniform influence area of 575
km2. Generally, rain gauging stations have smaller influence area towards the Bundelkhand plains, whereas, moderate
to large-size polygons are found in the central Bundelkhand upland and plateau region. Few polygons with a
comparable optimum size occur in the hilly region especially towards the headwater regions of the Betwa river. This
distorted distribution of rain gauges can be explained by the overall development in the region. Agricultural intensity
in the plains may be attributed to the establishment of more rain gauges there, reflecting the demand for accurate
measurement of rainfall in their vicinity. The rugged and inaccessible terrain in the middle course of the river, besides
the low level of development, has resulted in relatively large-size polygons. However, with the development of
irrigation in the region, new stations have been installed in Madhya Pradesh.

S. Rain- Influence Rainfall S. Rain-gauge Influence Rainfall

No. gauge area (km2) (mm) No. area (km2) (mm)
1 Talbahat 1806.86 837.5 30. Banda 787.86 969.5
2. Tikamgarh 1791.63 1001 31. Pathari 783.29 1253.4
3. Narhat 1648.50 997.9 32. Bijawar 772.32 1129.5
4. Sironj 1535.21 1011.2 33. Devanganj 691.01 1303.3
5. Mungaoli 1450.24 906.0 34 Raisen 658.12 1245.1
6. Ghoora Malehra 1359.18 1183.6 35. Barwasagar 616.40 796.8
7. Khurai 1292.49 1206.4 36. Moth 599.65 851.2
8. Berasia 1271.78 1047.9 37. Goharganj 585.64 1337.3
9. Basoda 1219.70 1141.1 38. Isagarh 558.84 908.5
10. Lalitpur 1202.65 981.5 39. Pachwara 526.25 766.5
11. Magarwara 1113.72 734.3 40. Belathal 517.73 860.9
12. Sarila 1103.06 810.7 41. Alipura 467.78 898.7
13. Sagar 1063.77 1229.4 42. Kulpahar 359.06 819.4
14. Bhilsa 1057.38 1159.7 43. Bilkisganj 347.49 1322
15. Begumganj 1017.18 1368.4 44. Hamirpur 336.52 865.9
16. Kurwai 1013.83 1101.0 45. Chiklod Kalan 319.16 1528.4
17. Chanderi 999.21 920.8 46. Nowgong 269.22 1027.7
18. Ashoknagar 983.07 863.5 47. Orai 265.26 734.9
19. Chandia Nallah 976.37 1284.4 48. Chatarpur 243.33 1067.5
20. Pichhore 958.10 876.5 49. Charkhari 226.28 885.1
21. Garaotha 920.33 870.3 50. Lugasi 172.37 869.9
22. Bhopal 910.28 1250.2 51. Silwani 135.52 1150.0
23. Mahroni 893.84 1025.5 52. Konch 112.07 759.8
24. Garrauli 891.40 873.3 53. Mahoba 88.32 927.1
25. Jhansi 821.97 917.6 54. Kalpi 87.40 859.9
26. Mau Ranipur 806.43 912.8 55. Karera 54.82 745.3
27. Mohammadgarh 800.34 1495.3 56. Bhander 27.41 767.1
28. Ghairatganj 798.82 1264.7 57. Maudaha 10.05 832.4
29. Rath 788.77 916.9 58. Sehore 5.79 1412.3
Table II: Mean Annual Rainfall and Influence Area of Rain Gauge by Theissen Polygon Method
(Source: Area Computed in ILWIS)

4.3. RAINFALL INTERPOLATION BY SURFACE FITTING: As the best areal estimates can be determined by
isohyetal method, therefore, a decision about search radius for points to be included within an isohyetal zone becomes
crucial parameter even for GIS software. Point-probability values provided significant clues in this direction (Table
III).
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 51

Lag Distance Probability of Finding Point Pairs (Rain-Gauging Stations)

class (meters) All Points 1Point 2 points 3 Points 4 Points 5 Points 6 Points
1 8200 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
2 9100 0.0004 0.0286 0.0000 0.0000 0.0000 0.0000 0.0000
3 10000 0.0012 0.0857 0.0000 0.0000 0.0000 0.0000 0.0000
4 11000 0.0021 0.1286 0.0143 0.0000 0.0000 0.0000 0.0000
5 12000 0.0021 0.1286 0.0143 0.0000 0.0000 0.0000 0.0000
6 13000 0.0029 0.1714 0.0286 0.0000 0.0000 0.0000 0.0000
7 15000 0.0041 0.2143 0.0571 0.0143 0.0000 0.0000 0.0000
8 16000 0.0050 0.2571 0.0714 0.0143 0.0000 0.0000 0.0000
9 18000 0.0062 0.3000 0.1143 0.0143 0.0000 0.0000 0.0000
10 20000 0.0070 0.3429 0.1286 0.0143 0.0000 0.0000 0.0000
11 22000 0.0099 0.4714 0.1714 0.0429 0.0000 0.0000 0.0000
12 24000 0.0145 0.5857 0.2571 0.1286 0.0286 0.0000 0.0000
13 27000 0.0186 0.6286 0.3714 0.1714 0.0571 0.0571 0.0000
14 30000 0.0236 0.7286 0.4571 0.2286 0.1286 0.0714 0.0143
15 33000 0.0306 0.8571 0.5857 0.3143 0.1857 0.1143 0.0429
16 36000 0.0385 0.9143 0.7286 0.4714 0.2714 0.1429 0.0857
17 39000 0.0464 0.9429 0.8143 0.6000 0.3714 0.2571 0.1143
18 43000 0.0555 0.9857 0.8571 0.7000 0.5143 0.3286 0.2000
19 47000 0.0654 0.9857 0.9143 0.7429 0.6143 0.4571 0.3143
20 51000 0.0841 0.9857 0.9571 0.8714 0.8143 0.6429 0.4571

21 56000 0.0994 1.0000 0.9714 0.9143 0.8714 0.8000 0.6429

22 62000 0.1226 1.0000 1.0000 0.9571 0.9143 0.8714 0.7857
23 68000 0.1416 1.0000 1.0000 0.9714 0.9286 0.9000 0.8857
24 75000 0.1681 1.0000 1.0000 1.0000 0.9857 0.9429 0.9286
25 82000 0.1909 1.0000 1.0000 1.0000 1.0000 0.9714 0.9286
26 91000 0.2327 1.0000 1.0000 1.0000 1.0000 1.0000 0.9714
27 100000 0.2687 1.0000 1.0000 1.0000 1.0000 1.0000 0.9857
28 110000 0.3114 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
29 120000 0.3565 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
30 130000 0.3975 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
31 150000 0.4791 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
32 160000 0.5122 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
33 180000 0.5847 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
34 200000 0.6476 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
35 220000 0.7101 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
36 240000 0.7615 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
37 270000 0.8306 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
38 300000 0.8841 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
39 330000 0.9317 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
40 360000 0.9598 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
41 390000 0.9793 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
42 430000 0.9934 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
43 470000 0.9996 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
44 510000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
Table III: Probabilities of Location of Rain-Gauging Stations in Betwa Catchment
(Source: Computed in ILWIS)
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 52

1.0

0.9

0.8

0.7
Probability of Points
0.6

0.5

0.4
Distance x Prob3Pnt
0.3 Distance x Prob2Pnt
Distance x Prob1Pnt
0.2 Distance x ProbAllPnt

0.1

0.0
0 100000 200000 300000 400000 500000
Distance (m)
Fig. IV: Probability of Finding a Location of Rain Gauge vs. Distance
(Source: Plotted in ILWIS based on Location Data)

A careful investigation of point-pattern graph based on all 70 points including nearby catchment area reveals
that the 100% probability to find at least one point lies at a distance of 56 km, 62 km for 2 points and 110 km for 6
points (Fig. IV). Therefore, search radius for any interpolation was used according to these limits. One point
probability based radius 56 km results into lesser correlation between modelled and experimental values than with 2
point probability radius. Hence, 62km was used for interpolation of rainfall data having a standard period of 30 year
with a network of 45 rain gauges.In order to obtain more comprehensive picture for normal annual rainfall data, the
data set having long-term record (>50 year) of rainfall with a network of only 25 rain gauge stations from the total was
used for comparison of methods (Table IVa). Isohyetal average obtained by contouring these precipitation values
revealed variation in estimates ranging from 962.11mm to 1033.95 mm (Fig. V a to g) A histogram-based analysis for
simplest measures like coefficient of variation (C v) reveals that although moving average linear decrease method
minimizes the overall variation but root mean square (RMS) error in rainfall estimates in this case is higher. Thus,
local variations remain hidden. Moving average inverse distance method minimizes the error near the station due to
close correspondence between the observed and expected values of rainfall in their vicinity. Trend surface (2 degree
polynomial method) and kriging provide good results of interpolation but large number of points are required for
computations and validations. Validation results for rest of the 11 stations having short-term record, however, show
somewhat different picture (Table IV b). Hence, error in both spatial and temporal variation are essential for optimum
results. However, utility of kriging method is proved in all three attempts in terms of reduction of overall variation in
areal means. Therefore, interpolation by different kriging methods was done to anlayse spatial variation in detail.

S.No. Method
Results using stations Results using stations with
with 30 yr record long term record (>50 yr)
Mean rainfall (Cv) Mean rainfall (Cv)
(mm) (%) (mm) (%)
1 Arithmetic average 969.53 16.75 981.00 16.84
2 Theissen polygon 953.72 15.02 962.11 13.90
3 Moving average (inverse distance) 1018.54 12.77 1033.17 10.76
4 Moving average (linear decrease) 1015.80 12.05 1032.96 9.59
5 Moving surface (2degree parabolic) 1019.96 14.28 1028.59 13.92
6 Moving surface (plane) 1023.73 13.12 1029.49 13.37
nd
7 Trend surface (2 degree polynomial) 1013.99 11.42 1030.21 12.66
8 Kriging (ordinary) 1021.30 12.08 1033.95 11.65
9 Kriging (simple) 1021.30 12.55 1032.91 11.72
Table IVa: Mean Rainfall and Variability
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 53

S.No. Method (Cv)(%) RMS R2

1 Moving average (inverse distance) 12.32 33.44 0.7425
2 Moving average (linear decrease) 12.15 37.24 0.6807
3 Moving surface (2degree parabolic) 12.63 32.28 0.7600
4 Moving surface (plane) 11.94 33.91 0.7353
5 Trend surface (linear) 10.73 36.96 0.6855
6 Trend surface (2nd degree polynomial) 11.04 37.43 0.6744
7 Kriging (ordinary) 11.62 35.34 0.7123
8 Kriging (simple) 11.77 35.43 0.7110
Table IV b: A Comparison of Validation Results

Table IV (a & b): Mean Rainfall of Betwa Catchment estimated using different Interpolation Methods
(Source: Computed in ILWIS based on rainfall data from i) Government of India, Indian Meteorological Department ,1999
Weekly Rainfall Probabilities. Vol. I and II; .ii) Climate of Madhya Pradesh, 1983; iii) Climate of Uttar Pradesh, 1989)

Fig. V a
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 54

Fig. V b Fig. V c

Fig. V d Fig. V e
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 55

Fig. V f Fig. V g

Fig.V (a to g): A Comparative View of Spatial Interpolation Methods for Rainfall

(Source: Author using ILWIS map computations based on Data cited in Table IV)
Further, the search radius was used to find spatial correlation among the pair of points in each distance class
for interpolation by Kriging method. It can be seen from Table V (a) that positive correlation (Moran’s I value) exists
till the 3rd distance class at 124 km. As the observed semi-variogram resembles the shape of Gaussian curve adaped
from the data (Fig. VI), therefore, modelled values of semi-variogram were obtained for different parameters that
resulted into good fit (r = 0.9895). Respective values of parameters to be used for kriging interpolation were as
follows:
Nugget = 3959.90; Sill = 104095.72; Range = 287761.50 m

S. Distance Number Moran’s Geary’s Average Semi-Variogram Semi-Variogram R2

No. (m) of Pairs I C Lag (m) Values Modelled Values

1 0 25. 0.834 0.11 22238.5 3959.90 0 0.9895

2 62000 290 0.632 0.28 65161.6 10091.81 8503.487 0.9895
3 124000 312 0.313 0.54 123970.4 19160.76 20934.272 0.9895
4 186000 284 -0.099 1.02 185002.5 36625.71 38165.007 0.9895
5 248000 198 -0.476 1.51 245916.8 53999.40 56461.419 0.9895
6 310000 130 -0.827 2.04 307409.9 72819.80 72732.683 0.9895
7 372000 60 -1.091 2.42 368291.5 86733.05 85277.319 0.9895
8 434000 24 -1.238 2.8 424144 100135.82 93805.762 0.9895
9 496000 3 -0.961 2.58 475720.6 92371.78 98968.159 0.9895
Table V (a): Omni directional Spatial Correlation Results
(Source: Author using ILWIS based on rainfall data cited in Table IV)
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 56

Distance vs. Semi-Variance

120000

100000

80000
Semi-Variance

60000

40000
Gaussian Model
Exponential Model
Spherical Model
Rational Quadratic Model
20000 WaveModel
Power Model
Distance x SemiVar
GaussianModel(Adapted)
0
0 50000 100000 150000 200000 250000 300000 350000 400000 450000
Distance (m)

Fig. VI: Semi-Variogram of Omni directional Spatial Correlation

(Source: Author using ILWIS Semi-variogram Plot based on data from Table Va)

Both simple and ordinary kriging methods using variogram operation results provide better estimates (Fig. VII
a & b). A localised trend in rainfall pattern can be observed from universal kriging with linear and quadratic surface
within a search radius of upto 16 points (Fig. VII c & d).

Fig. VII (a) Fig. VII (b)

Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 57

Fig. VII (c) Fig.VII (d)

Fig. VII (a to d): A Comparative View of Kriging-based (Omni Directional) Spatial Interpolation of Rainfall
(Source: Author using ILWIS map computations)

As variogram surface reflects anisotropy (Fig. VIII), therefore bidirectional spatial correlation results are
examined in Figure IX. Semi-variogram values in both the directions are given in Table V (b).

Fig. VIII: Anisotropic Variogram Model for Bidirectional Spatial Correlation

(Source: Author using ILWIS)
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 58

Distance x SemiVar1

40000

30000
SemiVar1

20000

Gaussian Model2
10000 Gaussian Model1
Distance x SemiVar2
Distance x SemiVar1

0
0 25000 50000 75000 100000 125000 150000 175000 200000 225000
Distance

Fig.IX: Semi-Variogram Model Adopted for Bidirectional Spatial Correlation

(Source: Author using ILWIS Semi-variogram Plot)

Lag Distance Moran’s Geary’s Horizontal Direction (53.5o) Mirror Rotate (133.5o)
class (m) I C Average No. Semi- Average No. Semi-
Lag (m) of Variogram Lag (m) of Variogram
Pairs Values Pairs Values
1 0 0.894 0.14 8931.1 2 6219.05 ? 0 ?
2 20000 0.904 0.16 21395.7 19 6489.48 22295.5 23 8062.49
3 40000 0.714 0.21 40632.4 51 9547.63 41466.5 51 9806.59
4 60000 0.665 0.29 59769.9 62 11391.22 58725.8 56 15472.75
5 80000 0.508 0.37 80237.5 94 13689.64 80939.0 56 21839.18
6 100000 0.312 0.48 100436.7 82 18249.61 101137.5 40 29359.03
7 120000 0.315 0.57 120270.3 92 17198.58 119326.5 36 47778.77
8 140000 0.158 0.70 140435.5 79 22143.77 140247.9 34 53428.59
9 160000 -0.166 0.98 160692.5 86 34323.36 160545.0 19 90475.45
10 180000 -0.086 0.86 179124.1 81 34472.81 180217.9 11 73195.59
11 200000 -0.342 1.45 200649.2 83 57813.22 199869.4 8 147340.80
12 220000 -0.387 1.43 219860.1 61 59381.07 222399.0 5 127905.30
13 240000 -0.509 1.61 240380.1 65 73000.65 ? 0 ?
14 260000 -0.581 1.54 259558.1 57 68955.37 251635.1 1 ?
15 280000 -0.922 2.27 279955.5 57 102599.10 ? 0 ?
16 300000 -1.042 2.31 299066.5 47 104392.90 ? 0 ?
17 320000 -0.999 2.38 319177.2 45 107419.50 ? 0 ?
18 340000 -1.045 2.32 338502.9 31 104810.60 ? 0 ?
19 360000 -1.194 2.59 360309.3 21 117078.10 ? 0 ?
20 380000 -1.224 2.55 379602.7 15 115511.10 ? 0 ?
21 400000 -1.420 2.97 398748.9 8 134308.70 ? 0 ?
22 420000 -1.300 2.72 415395.9 4 1229510.00 ? 0 ?
23 440000 -1.252 2.80 433633.3 2 1266450.00 ? 0 ?
? undefined pixels with no value calculated in that direction class

Table V (b): Bidirectional Spatial Correlation Results

(Source: Author using ILWIS based on rainfall data cited in Table IV)
The results of this bidirectional spatial correlation, when used in anisotropic kriging method, provide better estimates
(Fig. X a & b).
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 59

Fig. X (a) Fig.X (b)

Fig.X (a & b): Bi-Directional Effect in Spatial Interpolation of Rainfall

(Source: Author using ILWIS map computations based on Fig. VIII and IX)
Finally, error maps have been drawn for each type of kriging method that provide scientific guide for planning
(Fig. XI a to d).

Fig. XI (a) Fig. XI (b)

Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 60

Fig. XI (c) Fig. XI (d)

Fig.XI (a to d): A Comparative View of Kriging Residual Error
(Source: Author using ILWIS map computations based on Fig. X and Fig, VII)

Among error map of all types of kriging methods anisotropic method can be used for knowing the number of
stations as per optimum spatial variability. The error map obtained from simple kriging operation results into lowest
Cv. Based on error map, 95% confidence level map was drawn. It demarcates the region of 1000 mm rainfall till the
southern limit of Bundelkhand plain and a broad zone of transition lies between lower and higher limits, the latter
limit found towards the plateau region. However, the utility of kriging method remains limited because of different
search radii and several parameters used in semi-variogram model.

5. CONCLUSION: The method used for spatial estimates also affects the results as can be seen from differing results
of mean basin rainfall. Corresponding zone of influence of each station would also vary if area based computations are
used. Three interpolation methods - ordinary kriging, trend surface (2nd degree polynomial) and moving average
method with inverse distance weight function provided better spatial estimates (within 10-12% of coefficient of
variation) of mean catchment rainfall figures. Although, there exists no major difference in the results obtained from
these various interpolation methods but ordinary kriging method was found to be best. Kriging methods not only
considered spatial correlation of rainfall but also provided error map estimates with a maximum range of 15%
deviation from their mean. Hence, mean catchment rainfall derived from arithmetic average value of rainfall recorded
at stations and even linear interpolation methods like Theissen polygons can no longer be considered as easiest
substitute for spatially variable phenomenon. Among deterministic spatial interpolation methods - moving average
using inverse distance weight function can provide satisfactory result for less spatially correlated rainfall values in the
region. However, Kriging methods need large number of data stations in network for selection of parameters and their
calibration.

6. ACKNOWLEDGEMENTS: Author is grateful to Prof. Baleshwar Thakur (Retd.), Department of Geography,

University of Delhi for supervision of the work as part of Ph.D. Thesis (2007). The financial support of UGC in the
form of minor research project was crucial in both initial and completion stage of the work. Help from scientists at
Indian Institute of Remote Sensing as part of training sponsored by Department of Science and Technology was of
immense importance in clarification and understanding of the concepts related to GIS work. Acknowledgements are
also due to Officials of Indian Meteorological Department in providing access to data.

7. REFERENCES:
1. Jones, J.A.A., 1997, Global Hydrology: Processes, Resources and Environmental Management. Harlow, Essex: Addison
Wesley Longman Ltd. p. 120.
2. Bras, R.L. and Colon, R., 1978, Time-averaged areal mean of precipitation estimation and network design, Water Resource
Research, Vol. 14(5), pp. 878-88.
Ahlawat, R./ International Journal of Advancement in Remote Sensing, GIS and Geography, Vol.2, No. 1 61

3. Krige, D.G., 1966, Two-dimensional weighted moving average trend surfaces for ore-valuation, Proceedings of
Symposium on Mathematical Statistics and Computer Applications in Ore Valuation. South African Inst. Mining and
Metallurgy, Johannesburg, pp. 13-38.
4. Vagulin, F., 1999, Use of variograms to represent spatial uncertainty of geographic linear features. In: Kim Lowell and
Annick Jaton (eds.) Spatial Accuracy Assessment: Land Information Uncertainty in Natural Resources. Michigan: Ann
Arbor Press, pp. 157-64.
5. Goovaerts, P., 2000, Geostatistical approaches for incorporating elevation into the spatial interpolation of rainfall, Journal
of Hydrology, 228, pp. 113–129.
6. Bora, A.K. and Goswami, D.C., 1993, Designing on optimum rain-gauge network for estimating mean basin rainfall in the
Dhansiri basin, North-East India, Geographical Review of India, Vol. 55(3), pp. 18-24.
7. Prasad, R., 1996, Monitoring and Forecasting of Hydrological Parameters in the Mahanadi River Basin. Unpublished
Ph.D. Thesis. Delhi: University of Delhi, Department of Geography.
8. Saxena, .P., 1968, Bundelkhand - A case study in hydrography and water resources, Transactions of the Indian Council of
Geographers, Vol. 5, pp. 133-139.
9. Ahlawat, R., 1999, Water Resource Potential and Utilisation in the Bundelkhand Region. Unpublished M. Phil
Dissertation. University of Delhi: Department of Geography.
10. Ahlawat, R., 2007, Hydrological Data Network Analysis for Water Resource Planning in Lower Yamuna Basin: A Case
Study of Betwa River Catchment. Unpublished Ph.D. Thesis, Department of Geography, University of Delhi.