|Method and Results
|VEGETATION MAPS : HARD OR SOFT
With hard classification, each pixel
is attributed to the (vegetation) class it most closely resembles.
The result is a single map with a limited number of vegetation classes.
The hard classification technique is typically applied to high-resolution
images, such as SPOT-XS and Landsat-TM.
SPOT XS image 30 November 1994
© CNES - Distribution Spot Image
|However, when it is used to classify
low-resolution images, such as SPOT VEGETATION, a problem arises.
These images are built up from pixels each of which corresponds to
an area of 1000 x 1000 meters (100 hectares) in the field! Experience
has shown that several vegetation types generally exist within an
area of this size. For example, one pixel can consist of: an area
with 20 hectares of open water, surrounded by a zone of marsh vegetation
covering 30 hectares, in the middle of a vast dry area of 50 hectares
(dry grass with dispersed bushes). This is called a 'mixed pixel'.
A hard classification might attribute this pixel to the class "dry
grass with dispersed bushes", because that is what covers the
largest part of the pixel. All the other information is simply lost.
The primary purpose of a soft classification technique is precisely
to retain all this information.
SPOT VEGETATION image 25 October 1998.
With the aid of these images, vegetation maps can be prepared and
processes (dehydration, etc.) monitored.
A soft classification does not make
a single vegetation map, but rather a single map per vegetation
type. The mixed pixel is "unmixed" into its components.
Each image point on these maps has a value between 0 and 100, whereby
the value 0 indicates that the vegetation type does not appear in
the pixel, while the value 100 indicates that the entire pixel is
covered by it. In most cases, however, the value will lie somewhere
between these extremes. These maps are called 'fraction images'.
In the literature,
several methods are described to apply this unmixing. They
can be based on neural networks, fuzzy logic, or statistics.
This research uses the last technique.
What follows is - unavoidably - a bit of mathematics. According
to the theory of the "linear mixing model", the
reflectance of a pixel is equal to the sum of the reflectances
of its parts (think of the open water, marsh vegetation, etc.
described just a moment ago). Moreover, the reflectance of
such a part is directly proportional to the surface covered
(think of the 20, 30 and 50 percent in the example just given).
We can put this into the formula above, where r represents
the total reflectance (observed by the satellite), fi the
fraction or percentage which is covered per vegetation type
i and pi the pure reflectance of vegetation type i (as observed
if the entire pixel were covered by this vegetation).
This equation is then
converted into a system of equations by applying them to each
band (for one pixel) or to each pixel (for one band). The
number of unknowns is equal to the number of vegetation types
to be distinguished, and thus is much smaller then the number
of equations. Such systems are called "over-determined"
and can be solved with multiple regression techniques. The
above-mentioned fraction images
are the final result.
|THE SATELLITE REGISTERS CHANGES
Two methods were evaluated for monitoring
wetlands: Post-classification comparison and Multi-temporal change
vector analysis. Both are capable of detecting the sometimes subtle
changes, but there are important differences between them.
Accurate, but complicated and time-consuming
comparison", two sets of fraction images are necessary, one
for each point in time to be evaluated. Changes can be mapped out
by noting, for each vegetation type, the difference between the
fraction images. This method offers two advantages: even subtle
changes are recognisable (slight shifts within the same class, not
merely shifts from one class to another) and the results are a direct
interpretation of the vegetation. The only disadvantage is that
one must have two sets of fraction images, which costs both imaging
material and computer time.
Multitemporal change vector analysis
Less accurate, but faster results
Vector analysis of changes at
different points in time does not have this last disadvantage. Only
two sets of low-resolution images are necessary. From them, one
first derives a parameter with which the condition of the vegetation
can be described, such as for example a vegetation index (NDVI).
For each pixel, a multidimensional condition vector is then defined,
with each band corresponding to one dimension. Changes are mapped
out by subtracting these vectors pixel by pixel. Hereby both the
direction and the magnitude of the vector can be used as change
|VEGETATION MAPS : HARD OR SOFT
|This technology was applied on
a number of SPOT VEGETATION ten-day synthesis products dating from
August 1998 to June 1999. A large number of combinations were looked
at, varying from mono- to multi-temporal with a maximum number of
The figure shows the calculated fraction images of the entire flood
plain of the Logone for the three classes. The fraction images of
open water, marsh vegetation and dry vegetation are shown here from
left to right. For each of the fraction images, the colour varies
from black (value 0) through grey to white (value 100). The image
of the dry vegetation clearly shows that the flood plain is sharply
delineated (white is dry vegetation). The Logone River is also clearly
visible: the white line running through the middle of the plain in
the fraction image of water. On the right, moreover, you can see the
Chari River. The central section of the plain, on both sides of the
Logone, is dominated by marsh vegetation. This is primarily a wild
type of rice which grows in deep water. At the outer edge of the plain
there is primarily open water, with some marsh vegetation at points.
Result of a soft classification
From the left to the right are shown fraction images of open water,
marsh vegetation and arid vegetation
|THE SATELLITE REGISTERS CHANGES
rained there !
To assess the latter method, two
NOAA-AVHRR time series were used, placed free of charge on the Internet
by the Global Land 1 Km Data Set project. They date from October to
January of 1993 and 1995, and thus coincide with the drying out period
of the wetlands. This phase is decisive for plant growth and thus
is interesting for monitoring purposes.
Differences in NDVI between 1993 and 1995. Blue hues
indicate weak NDVI values in 1993, while red hues
point at higher NDVI values for 1993.
The figure reproduces the total NDVI-differences. It is striking
that blue hues dominate in the northern half of the image, which
corresponds to lower NDVI values in 1995, while in the southern
part the opposite is true and red hues dominate. This indicates
that in the South extra rainfall probably kept the vegetation green
longer in 1995 than in 1993.
This is confirmed by the blue tint of the Maga Reservoir, where
the rain produced a higher water level and the NDVI values - which
are negative for open water - consequently are lower than in 1993.
However, the problem with this method appears to be the interpretation
of these changes on vegetation areas. For this it can be useful
- along with the multidimensional approach - to incorporate the
one-dimensional changes into the interpretation as well.
With the aid of remote detection, and particularly low-resolution
images, vast wetlands can be mapped and their evolution monitored
over time. With the results of such research, policy-makers are
better able to estimate the consequences of human interventions
(building dams, laying irrigation networks, etc.). They can then
take these forecasts into account when making future plans, so that
these ecologically and economically valuable areas are preserved.