1
Reservoir capacity estimation of the Singoor Reservoir, India,
using per-pixel and sub-pixel classification approaches.
Jeyakanthan, V.S.* and Sanjeevi, S@
*Scientist-D, Deltaic Regional Centre, National Institute of Hydrology, Kakinada, A.P
@ Professor, Department of Geology, Anna University, Chennai
Abstract
Traditional approaches of image classification such as maximum likelihood and the band
thresholding method, involve the per-pixel approach to delineate the water spread area of a
reservoir. One of the limitations of these approaches is that the pixels representing the reservoir
border, containing a mixture of water, soil and vegetation, are classified entirely as water,
thereby resulting in, inaccurate estimate of the water spread area. To compute the water spread
area accurately, the sub-pixel approach has been used in this study. The water spread areas
extracted using per-pixel and sub-pixel approaches from IRS-1D and RESOURCESAT satellite
image data were in turn used to quantify the capacity of the Singoor Reservoir, Andhra Pradesh,
India. The estimated capacity of the reservoir using the per-pixel and sub-pixel approaches was
727.75 Mm3 and 716.11 Mm
3, respectively. The validation shows that the sub-pixel approach
produced much less error (1.08%) than the per-pixel based approach (3.14%).
Key Words: Reservoir, water spread area, capacity estimation, sub-pixel approach.
1. Introduction
Natural processes, such as erosion in the catchment area, movement of sediment and its
deposition in various parts of the reservoir, require careful consideration in the planning of major
reservoir projects. The silt that is deposited at different levels reduces the storage capacity of the
reservoir (Smith and Pavelsky 2009, Sreenivasulu and Udayabaskar 2010). Reduction in the
storage capacity beyond a limit prevents the reservoir from fulfilling the purpose for which it is
designed. Periodic capacity surveys of the reservoir help to assess the rate of sedimentation and
reduction in storage capacity. Conventional techniques for the estimation of the capacity of a
reservoir, such as hydrographic survey and inflow-outflow approaches, are cumbersome, time
consuming and expensive, and they involve significant manpower. As an alternative to
conventional methods, the remote sensing technique provides cost- and time-effective estimation
of the live capacity of a reservoir (Sabastian 1995, Jain 2002). Multi-date satellite remote sensing
data provide information on elevation contours, in the form of water spread area, at different
water levels of a reservoir. The water spread area thus interpreted from the satellite data is used
as an input into a simple volume estimation formula to calculate the capacity of a reservoir. Such
work has been reported by Manavalan (1990) for the Malaprabha Reservoir in India, Jeyaseelan
and Thiruvengadachari (1997) for the Kuttiyadi Hydel Reservoir in India, Chandrasekar and
Jeyaseelan (2000) for the Tawa Reservoir in India, Jeyakanthan (2002) for the Poondi Reservoir
in India, Goel (2002) for the Bargi Reservoir in India, Jain (2002) for the Bhakra Reservoir in
India and Peng (2006) for the Fegman Reservoir in China.
2
To quantify the capacity of a reservoir, the only thematic information that has to be
extracted from the satellite data is the water spread area at different water levels of the reservoir
(Morris and Fan 1998, Peng 2006). Different approaches to delineate various thematic
information from the remote sensing digital data, such as maximum likelihood classification,
minimum distance to mean classification and the band threshold method, adopt the per-pixel
based methodology and assign a pixel to a single land cover type (Jensen 1996, Bastin 1997);
whereas in reality, a single pixel may contain more than one type of land cover (known as a
mixed pixel). Mixed pixels are common especially near the boundaries of two or more discrete
classes (Foody and Cox 1994, Shalan et al. 2003, Ibrahim et al. 2005). The boundary pixels of
the water spread area that are mixed in nature, representing soil or vegetation with moisture, are
classified as water pixels when a per-pixel based approach is applied, thereby producing an
inaccurate estimate of the water spread area. To accurately compute the water spread area to the
maximum possible extent, thereby reducing the error in the estimation of the capacity of a
reservoir, a sub-pixel or linear mixture model (LMM) approach has been chosen for classifying
the boundary pixels of the water spread area from different water levels of the Singoor Reservoir
located in the Andhra Pradesh state of India.
2. Study reservoir
The Singoor Reservoir (Figure 1) is a major irrigation project built across the river
Manjira, which is one of the tributaries of the Godavari River, the second largest river in the
Indian subcontinent. The project is located near Singoor village, Medak District, which is at a
distance of 100 km from Hyderabad, the capital city of Andhra Pradesh state. The total length of
the dam is 7.52 km, which includes a 327 m long overflow masonry dam in the river gorge
portion and an 81 m non-overflow masonry dam flanked on both sides by earthen embankments.
The Singoor project is also used as the water supply source for the twin cities of Hyderabad and
Secunderabad. The climate of the sub-basin is characterised by a hot summer and a mild winter.
The monsoon season begins early in the month of June and continues to the end of October. The
principal type of soil present in the catchment area, apart from the red soil, is black cotton mixed
soil. Due to various activities in the catchment area, the black cotton soil is easily eroded and an
enormous amount of silt is carried into the stream that drains into the Manjira River. The eroded
soil eventually is deposited into the Singoor Reservoir and drastically reduces its capacity.
Hence, estimation of the capacity is a problem when studying this reservoir.
2.1 Satellite data used
The image data used in this study were acquired by the Indian Remote Sensing (IRS)
satellites IRS-1D & RESOURCESAT/IRS-P6 (LISS-III sensors) that provide a spatial resolution
of 23.50m in four different spectral bands (0.52-0.59, 0.62-0.68, 0.77-0.86, 1.55-1.70μm). IRS
satellite image data was selected to study the reservoir, because the images from the different
satellites viz. IRS 1C, 1D and P6 can be utilized from the year 1996 onwards with the higher
temporal resolution. In addition, these sensors have similar spectral resolutions which are
mandatory for reservoir sedimentation studies. The different dates of the satellite data used and
the respective water level during the pass of the satellite over the reservoir are given in Table 1.
Reservoir water level data and the hydrographic survey details have been collected from the
Singoor Reservoir authority responsible for the maintenance and operation of the reservoir.
3
Table 1. Selected water levels and the corresponding date of satellite
pass over the Singoor Reservoir.
Sl.No. Date of
Satellite Pass
Reservoir Water
Level above
mean sea level
(m)
Satellite &
Sensor
Path/
Row
1. 12.11.2005 523.49 IRS 1D-LISS III 99/60
2. 06.10.2005 522.10 IRS P6-LISS III 99/60
3. 29.08.2005 517.10 IRS 1D-LISS III 99/60
4. 01.04.2005 514.92 IRS 1D-LISS III 99/60
5. 15.05.2005 513.71 IRS P6-LISS III 99/60
6. 15.06.2005 512.51 IRS 1D-LISS III 99/60
3. Methodology
The changes in the water spread could be accurately estimated by analysing the areal
spread of the reservoir at different elevations over a period of time using the satellite image data
(Morris and Fan 1998, Smith and Pavelsky 2009). Per-pixel and sub-pixel approaches have been
used in this study to extract the water spread area of the reservoir. Estimated water spread areas
were used in a simple volume estimation formula to compute the storage capacity of the
reservoir. Estimation of the water spread area and the computation of the capacity of the
reservoir are discussed in the following sections.
Figure 1. Location map of the Singoor Reservoir.
4
3.1 Geo-referencing of satellite data
In the IRS-1D and P6 (LISS-III) satellite data the reservoir water spread area was free
from clouds and noise for all of the six temporal images used. The image scene of 12th
November 2005 was geo-referenced with respect to a set of 1:50,000 scale survey of India
topographic maps. The geo-referencing was performed using polyconic projection and the
nearest neighbour re-sampling technique to create a geo-referenced image with a pixel size of 24
m x 24 m. Subsequently, the other satellite images corresponding to various water levels were
also registered with the geo-referenced image using the image to image registration technique. In
every image, 25 to 30 ground control points were used, which resulted in a root mean squared
error (RMSE) of 0.12 to 0.19 of a pixel.
3.2 Per-pixel based approach
Water reflects most of the visible wavelengths, but the energy at the near-infrared (NIR)
wavelength is almost absorbed by the water, thus providing a significant contrast between land
and water in the NIR images (Lillesand and Kiefer 2004). This contrast helps in extracting the
water spread area of the reservoir. Different procedures have been adopted by many researchers
(Chopra et al. 2001, Sheng et al. 2001, Toyra et al. 2001, Dechka et al. 2002, Jain et al. 2002, Xu
2005, Rathore 2006, Alsdorf 2007) for water body identification in wetland areas and reservoirs,
each adopting the per-pixel based approach. Among these procedures, the band threshold
approach is a relatively easy and valid method for identifying the water body. It has also been
suggested that this per-pixel based approach can give acceptable estimates of the area of the
water body if the NIR band is used (Goel 1996, Jain 2002). Therefore, in the per-pixel based
approach, the band threshold technique was adopted to extract the water pixels that correspond to
various water levels of the reservoir. The following model equation has been used in the image
processing software to delineate the water spread area of the reservoir. The adopted algorithm
states that:
if
PV-NIR > TL-NIR and (1)
PV-NIR < TH-NIR then
the pixel is in the water spread area, where PV-NIR is the pixel value in NIR band and TL-NIR and
TH-NIR are the lower and higher thresholds for the NIR band.
Because the absorption of electromagnetic radiation by water is at a maximum in the NIR
spectral region, the digital number (DN) of water pixels is considerably lower than that
corresponding to other land cover types. Even if the water depth is shallow, the increased
absorption in the NIR band will restrict the DN value to less than that of the green and red bands.
If the soil is exposed (possibly saturated) at the surface, the reflectance will be as per the
signature of the soil, which increases with wavelength in this spectral range. Thus, by following
this algorithm (Equation 1) water pixels that belong to a particular water level of the reservoir
were extracted. The extracted water spread areas are shown in Figure 2.
5
Figure 2. Satellite data pertaining to six different water levels and the corresponding
extracted water spread area of the Singoor Reservoir based on the per-pixel approach.
Reservoir No. A B C D E F
Water level (m) 523.49 522.10 517.10 514.92 513.71 512.51
Extracted Area (Mm2) 138.37 110.96 52.01 22.36 17.28 10.32
A B C
D F E
6
3.3 Sub-pixel based approach
The sub-pixel classifier uses the linear unmixing technique that allows for the
identification of the “material of interest” and the determination of its “material part fraction” or
cover percentage within a pixel.
Linear spectral unmixing is a perfect approximation for calculating the abundance or
fraction of an end-member in an image pixel. The LMM classification technique attempts to
estimate the proportions of specific classes that occur within each pixel using the linear mixing
approach (Foody 1996, Aplin and Atkinson 2001, Min Xu 2005). In this study, the reservoir
water spread area was estimated using the linear spectral unmixing approach.
The basic assumption of the LMM is that the measured reflectance of a pixel is the linear
sum of the reflectance of the components that make up the pixel. The basic hypothesis is that the
image spectra are the result of mixtures of surface materials, shade and clouds and that each of
these components is linearly independent from the others (Bosdogianni et al. 1997, Atkinson
1997, Robert et al. 1998). Linear unmixing also assumes that all of the materials within the
image have sufficient spectral contrast to allow their separation. In a soft classification, the
estimated variables (the fractions or proportions of each land cover class) are continuous, ranging
from 0 to 100 percent coverage within a pixel. Settle and Drake (1993) and Foody and Cox
(1994) proposed a mathematical expression for linear spectral unmixing. The theory behind this
is that a series of end-members present within a pixel contribute to the overall spectral signature
of that pixel. Hence, the spectral signature of a pixel would be derived from the sum of the
products of the single spectrum of each end-member it contains, each weighted by a fraction,
plus a residue as explained by the following mathematical model:
Ri = ∑fk Rik + Ei (2)
where ∑ fk = 1 (3)
and 0 ≤ fk ≤ 1 (4)
i = 1, . . ., m (number of spectral bands)
k = 1,. . ., n (number of end-members)
Ri = Spectral reflectance of band i of a pixel which contains one or more end-members
f k = Proportion of end-member k within the pixel
Rik = Known spectral reflectance of end-member k within the pixel in band i
Ei = Error for band i (Difference between the observed pixel reflectance Ri and the
reflectance of that pixel computed from the model).
Equations 2 and 3 introduce the constraints that the sums of the fractions are equal to one and
they are non-negative. To solve for fk, the following conditions must be satisfied: (i) selected
end-members should be independent of each other, (ii) the number of end-members should be
less than or equal to the spectral bands used, and (iii) selected spectral bands should not be
highly correlated.
In this study, linear spectral unmixing is adopted based on the equations described below to
segregate the actual information within a pixel of an image:
7
R1 = Fwater * R1water + FVeg* R1Veg + FSoil*R1Soil + ε1
R2 = Fwater * R2water + FVeg* R2Veg + FSoil*R2Soil + ε2 (5)
R3 = Fwater * R3water + FVeg* R3Veg + FSoil*R3Soil + ε3
where,
R1, R2 and R3 represent the signal recorded by the satellite (DN values) in the green, red and
NIR bands of the LISS-III sensor.
Fwater , FVeg and FSoil are the fraction of the pixel covered by water, vegetation and soil.
R1water, R2water and R3water represent the DN of water in each of the three spectral bands (water
end-members).
R1Veg, R2Veg and R3Veg represent the DN of vegetation in each of the three spectral bands
(vegetation end-members).
R1Soil, R2Soil and R3Soil represent the DN of soil in each of the three spectral Bands (soil end-
members).
ε1, ε2 and ε3 are the error components of band 1, 2 and 3.
The system of linear equations shown above can be solved by a least squares solution that
minimises the sum of squares of errors. To maintain the consistency between the per-pixel and
sub-pixel processes DN values were used as input parameter in the equation 1 & 5.
The sub-pixel based approach was applied to determine the proportion or fraction of the
water class that exits in the peripheral pixels of the reservoir. The first step executed in the sub-
pixel approach was the selection of the pure pixels (known as end-members) belonging to a
specific class. In general, the border pixels may contain any combination and proportion of
water, vegetation and soil classes. Hence, these three classes were chosen to represent the end-
members. The scatter plot method was used to identify the end-members. The locations of the
end-members in the image data were identified from the extremes of the scatter plot. The
identified end-member spectra were supplied as input to the LMM approach. The output of the
model run contains three images labelled as water-, soil- and vegetation- fraction images. A
description of the fraction images is given in section 4.2.
3.4 Computation of volume between successive water levels
Traditionally the reservoir volume between two consecutive reservoir water levels was computed
using the prismoidal formula, the Simpson formula or the trapezoidal formula (Patra 2001). Of
these, the trapezoidal formula has been most widely used for the computation of volume (Rao et
al. 1985, Goel and Jain 1996, Morris and Fan 1998, Rathore 2006). The water spread area
estimated using the per-pixel and sub-pixel approaches were separately used as an input to the
volume estimation formula to determine the volume at different water levels of the reservoir. In
8
this study the volume between two consecutive reservoir water levels was computed using the
following trapezoidal formula:
V = H *(A1 + A2 + √A1*A2)/3 (6)
where V is the volume between two consecutive water levels, A1 and A2 are the water spread
areas at the reservoir water levels 1 and 2 respectively and H is the difference between these two
water levels.
3.5 Computation of storage capacity of the reservoir
The volumes computed (using equation 6) between different water levels (i.e., from minimum
draw down level (MDDL) to full reservoir level (FRL)) were added together to calculate the
cumulative or storage capacity of the reservoir.
4. Results and discussion
4.1 Computation of reservoir capacity by the per-pixel approach
Six different water levels varying from 512.51 m to 523.49 m of the reservoir were selected
based on the availability of cloud free satellite data to estimate the water spread of the reservoir
(the impoundment MDDL and FRL of the reservoir are 510.6 m and 523.6 m respectively). To
extract the water pixels from the images using the per-pixel approach, the algorithm presented in
Equation 1 was used. This algorithm requires separate minimum and maximum threshold DNs
from the NIR band for the six satellite images used in the study. With the help of the algorithm,
the pixels which contain a DN between the given minimum and maximum threshold values were
labelled as water pixels. Basic statistical parameters of an image that could be obtained using the
image processing software were utilised to determine the minimum pixel values of the water
spread area. The locations of these pixels in the water spread area were also verified, and showed
that the pixels with low DN were located in the deeper and central portion of the water spread
area of the reservoir. The lower (14) and higher (41) DN values can be attributed to the
irradiance of the water-body during the winter and summer seasons of the study area. The
extracted minimum DN were 28, 41, 25, 21, 21 and 14 for the images pertaining to the months of
April, June, July, August, October and November, respectively. The pixel values of 14 and 41
pertain to the winter (November) and summer (June) seasons of the study area. The analysis of
DN values of the water body show that the pixel value increases towards the periphery of the
water body and the border pixels have the maximum DN. In selecting the maximum value of the
water spread area, one must examine the pixels along the periphery of the reservoir. However, it
is not an easy task to select a maximum threshold value along the border of the reservoir area.
For example, at one location of the border area, one may be satisfied with a pixel value of 35, but
in another location it may change to a DN value of 43. Thus, the differences in pixel values over
a large range consume enormous time and effort to obtain a conclusive maximum threshold for
the DN to extract the water spread area. The extracted maximum pixel values were 36, 49, 44,
43, 49 and 37 for the images pertaining to the months of April, June, July, August, October and
November, respectively.
9
Table 2. Capacity estimation of the Singoor Reservoir using per-pixel and sub-pixel based
approaches.
Date of
Satellite
Pass
Reservoir
Elevation
above
m.s.l (m)
Water spread
area – Per-
pixel
approach
(Mm2)
Water spread
area – Sub-
pixel approach
(Mm2)
Cumulative
Volume –
Per-pixel
approach
(Mm3)
Cumulative
Volume –Sub-
pixel approach
(Mm3)
12.11.05 523.49 138.37 136.37 727.75 716.11
06.10.05 522.10 110.96 110.36 554.82 544.95
29.08.05 517.10 52.01 50.51 156.59 152.40
01.04.05 514.92 22.36 21.59 77.67 76.01
15.05.05 513.71 17.28 16.46 53.85 53.06
15.06.05 512.51 10.32 9.82 37.47 37.47
The total number of water pixels that were extracted was multiplied by the area (24 m x
24 m) of a single pixel to compute the water spread area. The same technique was adopted to
convert the extracted water pixels into the water spread area in all six images used in this study.
The water spread area thus estimated in each image using the per-pixel approach has been
used as an input to the trapezoidal formula (Equation 6) to calculate the consecutive volumes of
the reservoir. The storage capacity between the bed level (500.17 m) of the reservoir and the
lowest observed water level (512.51 m) could not be estimated using remote sensing
methodology due to the unavailability of cloud free satellite data. Therefore, the capacity (37.47
Mm3) between these two levels was adopted from the 1997 elevation-capacity table, available
from the Dam Authority. Above the lowest observed level (512.51 m), the estimated capacities
between the consecutive water levels were added up to arrive at the cumulative capacity of the
reservoir at the maximum observed level (523.49 m). The estimated cumulative capacity of the
reservoir at a water level of 523.49 m (near FRL) using the per-pixel classification approach was
727.75 Mm3. Thus, the per-pixel approach was adopted for all six images and the capacity was
estimated (Table 2). To compare the efficiency of this method with that of the sub-pixel
approach, another set of experiments was carried out, as described in the next section.
4.2 Computation of reservoir capacity by the sub-pixel approach
The fraction images (Figure 3 and 3a) generated using the sub-pixel approach described in the
Methodology section contain a wealth of information about the reservoir. Each fraction image
corresponds to a single type of land cover only. For example, the pixels in the water fraction
image provide information only on the proportion or amount of water contained in the pixel.
Likewise, the vegetation and soil fraction images provide information on the proportions of their
respective classes only. However, in this study the interest is only to determine the amount of
water present in the border pixels of the reservoir. The value of the pixels in the fraction image
10
Figure 3. (a) Feature space plot (NIR Vs RED),(b) End-member spectra of Soil, Water and Vegetation (c) Fraction
images obtained by spectral unmixing of image data pertaining to the highest water level (523.49m)
Figure 3a. (a) Feature space plot (NIR Vs RED, (b) End-member spectra of Soil, Water and Vegetation
(c) Fraction images obtained by spectral unmixing of image data pertaining to the
lowest water level (512.51m)
(Due to the recurring nature of the fraction images only figures pertaining to the highest and lowest water levels are
presented)
Dig
ital
Nu
mb
er
R
ED
NIR
Band
(b) (a)
NIR
R
ED
(a) (b)
Dig
ita
l N
um
ber
Band
11
ranges from 0 to 1. A pixel from the water fraction image having a value of 0 indicates that there
is no water at all in that pixel, whereas a pixel having a value of 0.3 indicates that 30% of the
area of the pixel is occupied by water while a pixel value of 1 indicates that 100% of the area of
the pixel is occupied by water (i.e., the pixel is fully occupied by water). Therefore, for a pixel
having a value of 0.7, the area of water occupied by that pixel is 403.2 m2 (0.7 x 24 m x 24 m).
The pixels representing the peripheral portion of the reservoir, which have a minimum
value of 0.1 in the water fraction image (i.e., a pixel with a minimum of 10% of its area
containing water) were isolated from the water-fraction image and the area covered by water in
these peripheral/border pixels was estimated. After examining the peripheral pixels it was
ascertained that none of the border pixels contain water spread area less than 10%. Hence, a
threshold value of 10% was selected and used for analysis. The number of pixels that contain
100% water was also determined. By summing the area occupied by these two types of pixels,
the total water spread area corresponding to a particular water level of the reservoir was
computed. This exercise was carried out for each of the six images used in the study. The water
spread area thus estimated was again used as an input to the trapezoidal formula to compute the
storage capacity or cumulative capacity of the Singoor Reservoir using the sub-pixel
classification approach.
It is worth mentioning that a pixel containing 65% water may be labelled as containing
100% water by the per-pixel approach. Thus, the water spread area is overestimated. Conversely,
if the pixel contains 40% water then the entire pixel is not considered for the water spread area
estimation. Hence, the water spread area is underestimated. Such errors due to overestimation or
underestimation do not occur in the sub-pixel approach. Thus, the sub-pixel approach reduces the
error imposed by the per-pixel approach. The estimated cumulative capacity of the reservoir at a
water level of 523.49 m (near FRL) using the sub-pixel approach was 716.11 Mm3. The capacity
estimated using the sub-pixel approach is given in Table 2. Elevation capacity curve of Singoor
Reservoir using per-pixel and sub-pixel approaches is given in Figure 4.
Figure 4. Elevation-Capacity curve of Singoor Reservoir using per-pixel and sub-pixel
approaches.
12
4.3 Validation of the per-pixel and sub-pixel approaches
Several investigators (Quaramby et al. 1992, Oleson et al. 1994, Haertel et al. 2004,
Foody 2007) have shown that the recovery of sub-pixel information from medium resolution
data is feasible, and this information can be directly compared to that obtained at higher scales.
In line with the above findings, it was decided to validate the result of the sub-pixel classification
approach (which was carried out using the 24 m resolution image data) using high resolution
image data with a spatial resolution of 5 m (resampled IRS 1C-PAN). However, concurrent high
resolution data was not available for any of the six images used in the study. Therefore, three
new sets of image data were procured in such a way that for a particular water level of the
reservoir, both the 24 m and 5 m resolution image data were available. The band threshold and
sub-pixel classification techniques were again carried out on the new 24 m resolution image data
sets. The high resolution PAN data were classified using the band threshold method. The results
of these two experiments are given in Table 3. By analysing the results for all three sets of
images, it was ascertained that the band threshold method applied to the 24 m resolution image
data
Table 3. Validation of the per-pixel and sub-pixel approaches.
Satellite/Sensor Reservoir
water level
above m.s.l
(m)
IRS-P6 / LISS-III (24 m) IRS-1C/PAN
(5m)
Date of Satellite Pass Water spread area
Per-Pixel (Mm2)
Water spread area
Sub-Pixel (Mm2)
Water spread area
Per-pixel (Mm2)
03 Feb 2006 (Validation-1) 522.68 123.73 118.69 119.82
23 Mar 2006 (Validation-2) 521.71 104.89 106.73 107.49
15 Jan 2005 (Validation-3) 516.55 42.04 39.68 40.32
Table 4. Percentage (%) of error between the per-pixel and sub-pixel based approaches.
No. of
Validation
Processing Approach
Per-pixel Sub-pixel
Validation-1 3.26% 0.94%
Validation-2 2.42% 0.71%
Validation-3 3.75% 1.58%
Average Error 3.14% 1.08%
overestimates whereas the sub-pixel method underestimates the water spread area when
compared to the high resolution (5 m) data. To select the best approach the percentage of error
between both these approaches were calculated. For example the percentage of error for
validation 2 was calculated as follow: ((104.89 – 107.49)/107.49) x 100 = 2.42% and ((106.73 –
107.49)/107.49) x 100 = 0.71%. Results of percentage of error between per-pixel and sub-pixel
approaches are given in Table 4. Analysis of Table 4 reveals that the sub-pixel approach
13
produced much less error (1.08%) than the band threshold/per-pixel approach (3.14%). This
shows that the sub-pixel based approach can be applied to estimate the capacity of the Singoor
Reservoir with higher accuracy than the per-pixel approach.
5. Conclusion
High spatial-resolution image data enables accurate mapping of terrain features. The use
of high spatial resolution satellite image data, however, is constrained by factors such as cost and
the smaller area covered by the sensor. Hence, in hydrological applications estimating the water
spread area may be difficult because a reservoir may not be imaged in a single pass of the
satellite and atmospheric conditions would be different from path to path (Hung and Wu 2005).
An alternative method to overcome such constraints is the use of the sub-pixel based approach.
The simplest methodology for such an approach is the linear mixture model, which has been
demonstrated in this study to accurately estimate the capacity of the Singoor Reservoir in India.
In this study both the per-pixel and sub-pixel approaches have been performed to extract
the water spread area of the reservoir using medium resolution multi-spectral image data and the
results were validated using high resolution panchromatic image data. The validation shows that
the application of the sub-pixel approach produced much less error (1.08%) than the per-pixel
(3.14%) based approach. The relatively lower error shown by the sub-pixel approach presents a
potential use of the technology for the estimation of other reservoirs with acceptable accuracy.
Although the sub-pixel approach was found to be a better alternative than the per-pixel
approach, there are certain limitations to the method, such as the fact that the spatial locations of
the various fractions within a pixel are unknown. In addition, the sub-pixel classifier produces
more accurate results only with hyperspectral images. Hence, the use of hyperspectral image data
with higher spatial resolution would have yielded better results.
References
Alsdorf, D.E., Rodriguez, E., Lettenmaier, D.P., 2007. Measuring surface water from space. Reviews of
Geophysics, 45 (2), 1–24.
Aplin, P., and Atkinson, P.M., 2001. Sub-pixel land cover mapping for per-field classification.
International Journal of Remote Sensing, 22 (14), 2853-2858.
Atkinson, P.M., 1997. Mapping sub-pixel proportional land cover with AVHRR imagery. International
Journal of Remote Sensing, 18 (4), 917-935.
Bajjouk, T., Populus, J., and Guillaumont, B., 1998. Quantification of sub- pixel cover fractions using
PCA and a linear programming method, Application to the Coastal Zone of Riscoff (France). Remote
Sensing of Environment, 64, 153-165.
Bastin, L., 1997. Comparison of fuzzy c-means classification, linear mixture modelling and MLC
probabilities as tools for unmixing coarse pixels. International Journal of Remote Sensing, 18 (17), 3629-
3648.
Bosdogianni, P., Petrou, M., and Kittler, J., 1997. Mixed pixel classification with robust statistics. IEEE
Transactions on Geoscience and Remote Sensing, 35 (3), 551-559.
14
Chandrasekar, K and Jeyaseelan, A.T., (2000). Sedimentation of Tawa reservoir through remote sensing.
Tenth National Symposium on Hydrology, 18 and 19 July, 2000, New Delhi, 509-522.
Chopra, R., Verma, V.K. and Sharma, P.K., 2001. Mapping, monitoring and conservation of Harike
wetland ecosystem, Punjab, India through remote sensing. International Journal of Remote Sensing, 22
(1), 89-98.
Dechka, J.A., Franklin, S.E., Watmough, M.D., Bennett, R.P. and Ingstrup, D.W., 2002. Classification of
wetland habitat and vegetation communities using multitemporal IKONOS imagery in southern
Saskatchewan. Canadian Journal of Remote Sensing, 28 (5), 679–685.
Foody, G. M., and Cox, D. P., 1994. Sub-pixel land cover composition estimation using a linear mixture
model and fuzzy membership functions. International Journal of Remote Sensing, l5 (3), 619-631.
Foody, G.M., 1996. Approaches for the production and evaluation of fuzzy land cover classifications
from remotely sensed data. International Journal of Remote Sensing, 17 (7), 1317-1340.
Foody, G.M., 2007. Variability in soft classification prediction and its implications for sub-pixel scale
change detection and super resolution mapping. Photogrammetric Engineering and Remote Sensing, 73,
923-933.
Goel, M. K., and Jain, S. K., 1996. Evaluation of reservoir sedimentation using multi-temporal IRS-1A
LISS II data. Asian Pacific Remote Sensing & GIS Journal, 8 (2), 39-43.
Goel, M.K., Jain, S.K., and Agarwal, P.K., 2002. Assessment of sediment deposition rate in Bargi
Reservoir using digital image processing, Hydrological Sciences Journal, 47(S), S81-S92.
Haertel, V, Shimabukaro, Y.E., and Almedia-Filho, R., 2004. Fraction images in multitemporal change
detection. International Journal of Remote Sensing, 25 (23), 5473-5489.
Hung, M.C., and Wu. Y.H., 2005. Mapping and visualizing the Great Salt Lake landscape dynamics using
multi-temporal satellite images-1972-1996. International Journal of Remote Sensing, 26 (9), 1815-1834.
Ibrahim, M. A., Arora, M. K., and Ghosh, S. K., 2005. Estimating and Accommodating uncertainties
through soft classification of remote sensing data, International Journal of Remote Sensing, 26 (14),
2995-3007.
Jain, S.K., Singh, P., and Seth, S.M., 2002. Assessment of sedimentation in Bhakra Reservoir in the
western Himalayan region using remotely sensed data. Hydrological Sciences Journal, 47 (2), 203-212.
Jensen R.J., 1996. Introductory Digital Image Processing : A remote sensing perspective, 2nd edition,
(London: Prentice Hall International Ltd.,).
Jeyakanthan, V.S., Sreenivasulu, V., Rao, Y.R.S., Ramasastri, K.S., (2002). Reservoir capacity estimation
using satellite data. IAPRS & SIS, Vol.34, Resources and Environmental Monitoring, Hyderabad, India,
863-866.
Jeyaseelan, A.T., and Thiruvengadachari, S., (1997). Sedimentation survey of Kuttiyadi Hydel reservoir through
remote sensing method. Proceedings of the Workshop on Remote Sensing and GIS applivation in Water Resources
and Engineering, 17-19, September 1997, Bangalore, India, 8-15.
15
Lillesand, T.M. and Kiefer, R.W., 1994. Remote Sensing and Image Interpretation, 3rd edition, (New
York: John Wiley & Sons).
Manavalan, P., Sathyanath, P., Sathyanarayn, M., and Raje Gowda, G. L., 1990. Capacity evaluation of the
Malaprabha reservoir using digital analysis of satellite data. Tech. Report no. RC: BO: WR: 001:90, Regional
Remote Sensing Service Centre, Bangalore and Karnataka Engineering Research Station, Krishnarajsagar. India.
Min Xu, Watanachaturaporn, P., Varshney, P. K., and Arora, M.K., 2005. Decision Tree Regression for
Soft Classification of Remote Sensing Data, Remote Sensing of Environment, 97 (3), 322-336.
Morris, G.L., and Fan Jiahua., 1998. Reservoir Sedimentation Handbook, (New York: McGraw-Hill
Book Co).
Patra, K.C., 2001. Hydrology and Water Resources Engineering, (New Delhi: Narosa Publishing House).
Rathore, D.S., Anju Choudhary, and Agarwal, P.K., 2006. Assessment of sedimentation in Harakud
Reservoir using digital remote sensing technique. Journal of the Indian Society of Remote Sensing, 34,
344-383.
Robert, D. A., Gardener, M., Church, R., Ustin, S., Scheer, G. and Green, R. O., 1998. Mapping
Chaparral in the Santa Monica Mountains using Multiple End member Spectral Mixture Analysis. Remote
Sensing of Environment, 65, 267-279.
Sabastin, M., Rao, P.P.N., Jayaraman. V., and Chandrasekhar, M.G., 1995. Reservoir storage loss
assessment and sedimentation modeling through remote sensing techniques. Proceedings of 6th Inter.
Symp. River sedimentation on Management of Sediment – Philosophy, Aims and Techniques, New Delhi,
India.
Settle, J.J., and Drake, N.A., 1993. Linear mixing and the estimation of ground cover proportions.
International Journal of Remote Sensing, 14 (6), 1159-1177.
Shalan, M. A., Arora, M. K., and Ghosh, S. K., 2003. An evaluation of fuzzy classifications from IRS 1C
LISS III data. International Journal of Remote Sensing, 23 (15), 3179 - 3186.
Sheng, Y., Gong, P., and Xiao, Q., 2001. Quantitative dynamic flood monitoring with NOAA AVHRR.
International Journal of Remote Sensing, 22 (9), 1709–1724.
Smith, L.C., and Pavelsky, T.M., 2009. Remote sensing of volumetric storage changes in lakes. Earth
Surface Process and Landforms, 34, 1353 – 1358.
Sreenivasulu, V and Udayabaskar, P., 2010. An integrated approach for prioritization of reservoir
catchment using remote sensing and geographic information system, Geocarto International, 25 (2),149-
168.
Toyra, J., Pietroniro, A. and Martz, L.W., 2001. Multisensor hydrologic assessment of a freshwater
wetland. Remote Sensing of Environment, 75 (2), 162–173.
Xu Hanqiu, 2005. Modification of normalized water index (NDWI) to enhance open water features in
remotely sensed imagery. International Journal of Remote Sensing, 27 (14), 3025 – 3033