In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
*
466
were then chosen: 1) enhanced image classification and 2) thresh
olding the normalized difference water index (NDWI; McFeeters,
1996; Ji et al„ 2009).
Image Classification. Traditional image classification is carried
out on a pixel-by-pixel basis. Although an increasing number
of studies show that region-based or object-based classification
tends to improve results significantly, it was not judged neces
sary in this particular case for two main reasons. First because
of the spectral nature of water being a smooth surface with small
variations (at least in the optical infrared) and secondly because
these methods usually offer little control over what is defined as
an object. Conversely, classification approaches such as Maxi
mum likelihood can produce posterior probability maps that can
be thereafter thresholded (hardened). The latter approach had the
advantage to require training data only for the object of interest
whereas classical classification procedures require all classes to
have been defined using training data. In this case we opted for
the posterior probability which can be simplified as the Gaussian
probability density of the ’’water class”. In simple nominal classi
fication, a pixel can be classified as pertaining to a particular class
even if its probability is low, as long as it is higher than for all the
other classes. By using a high threshold value (i.e. > 90%) to
attribute a water label to a pixel, we are able to use but a single
class and avoid having to gather training data for other objects or
surfaces.
NDWI threshold. Using the same logic as the normalized differ
ence vegetation index (NDVI) the normalized difference water
index (NDWI) was proposed by McFeeters (1996) as a means to
separate water from other surfaces (Eq. 3).
NDWI = p9reen ~? NIR (3)
Pgreen “F pNIR
where p gr een is the green reflectance (Landsat TM band 2: 0, 52—
0,60pm) and pnir is the near infrared reflectance (Landsat TM
band 4: 0,77 — 0,90pm). The NDWI varies between -1 and
1 and uses zero as the threshold between land (< 0) and water
(> 0). A number of variations were later proposed for NDWI.
In their article, Ji et al. (2009) compared a number of these vari
ations applied to Landsat, ASTER, SPOT and MODIS images.
They found that the modified NDWI (MNDWI) proposed by Xu
(2006) performed better (Eq. 4).
MNDWI = p9reen ~ PSWI . H (4)
Pgreen “F PSWIR
where pswiR is the reflectance in short wave infrared (Landsat
TM band 5: 1, 55 — 1,75p,m).
2.4 Validation and Statistical Testing
Two validation data sets were unsed for testing the performance
of the extraction of the lake contours from the interpolated Land
sat data which also involved our definition of the “water-land”
edge. First, the contours from the dry season image of 2006 were
compared against the contours extracted from a fusionned Ikonos
image (1 m) five days apart form the Landsat image. Secondly,
the four lakes of the VPSP (data from the larger lake outside the
park could not be acquired) were surveyed using a geodetic GPS
in kinetic mode to be compared with the contour from the Land
sat image (with a five days difference). Coordinates of the lake
contour were acquired at an interval of 15 meters with an approx
imate precision of 10 cm.
The validation was done by two complementary methods: 1) by
expressing the difference between the areas as a proportion of
the validated area ( y4reoi ~' 4 ° b ^ er ' i ' ed x 100); and 2) by overlap
ping the two contours (interpolated Landsat and validation data)
and dividing the overlap area (intersection) by the merged areas
(union) of both contours as illustrated in Figure 3.
Figure 3: Validation method for testing the accuracy of the lake
contours extracted from the interpolated Landsat images.
The statistical testing consists in establishing the strength of the
relationship between the areas of all six lakes and the AW of the
same period as the images. Although the response of the wa
ter level is not spontaneous, the trend should still be statistically
perceptible. Because the areas of the lakes are not normally dis
tributed, a regression was not recommended. Spearman’s cor
relation does not assume a normal distribution of the dependant
variable and was chosen instead. The correlation was also com
puted between the area of the lakes themselves as a mean to infer
a generalized trend.
3 RESULTS AND DISCUSSION
3.1 water balance
The AW was calculated for the period 1983 - 2009 using the
Thonthwaite method trimonthly (the year 1983 was added in or
der to feed the Available Water for the beginning of the 1984
budget). Figure 2 shows the annual budget averaged every five
years for the period along with the average budget for the whole
period (white line). Apart from the two first periods (1984-1989
and 1990-1994) which appear as exceptionally high and excep
tionally low respectively, the other periods do not show any trend
towards an increase or a decrease.
125
75 \
\A\
nA
25
— — 1984-1988 1
1990-1994 i
1995-1999 J
2000-2004 /
2005-2009 I
Average (1984-2009) /
-25/ A
; / / / 4
4
\ ■ '
! mi
-75
jf
-125
Figure 4: water balance averaged for every five years between
1984 and 2009 and overall average (white line).
3.2 Lake Contours Extraction and validation
The 50 selected Landsat images were geometrically corrected,
registered to a UTM grid, corrected for atmospheric interferences