|
Methodology
The methodology flowchart shows the different processing
steps that have been taken in order to be able to merge the entire data set
together. The processing has been divided into four main blocks:
- Classification
In the classification stage, a supervised classification
procedure was applied to each image using a maximum likelihood algorithm. The
selection of the training sites was aided with the results of a first
unsupervised classification, which helped to determine those classes
spectrally separable, and NDVI analysis used to determine areas of high biomas
and shadows in the images due to topographic effects. Also, the display of
false colour composites helped determining areas covered by vegetation. Images
were classified into 5 different groups: water, montane forest, grassland,
scrub/tree savannah and farmland.
- Geometric corrections
Due to the mountainous nature of the area (a volcanic cone
ranging from 800 to 3000 meters), important distortions due to relief
displacement were present in the images, especially in the aerial photographs.
For this reason, image orthorectification was required. For both aerial photos
and satellite images, a single frame orthorectification was carried out using
a DEM
derived from a Radarsat stereo-pair. Ground control points (GCP’s) were
obtained from a topographic map of the area and the entire dataset was
registered to the Universal Transverse Mercator (UTM) projection.
Image manipulation
The image manipulation stage was different for aerial
photos and satellite images. For aerial photographs, since the forest
boundaries were quite discernible, no other processing technique apart from
"heads up" on-screen digitising was necessary. The output of the digitising process is a vector layer that was converted into a raster image in
order to merge the entire data set together.
For the satellite images, after orthorectification, the
five categories in which the original images were classified were merged and
re-coded to create binary images of "forest" / "non
forest". These binary images showed pixels that were classified as forest
but located outside the forest boundaries. These mis-classified pixels were
masked out using the rasterised 1958 image (the earliest data available) as a
binary mask. This approach, therefore, assumed that no regeneration had taken
place outside the forest limits defined in 1958.
-
GIS integration
At this point we have a set of five binary images showing
the forest extension at different dates. These images were then combined in a
GIS environment. Image couples of successive dates were analysed in order to
obtain information about change trajectories, (i.e. deforestation or
regeneration patterns). For this purpose, the images in each of the four
combinations were re-coded. The first date image kept the 0/1 coding for
"forest/ non forest" and the second date image was then re-coded
into 0/2 for "forest/ non forest". Consequently, when images were
combined (by addition), the output image presented four different values (0-3)
related to number of times each pixel of the final output appeared in the
input images:
0 for "non forest"
1 for those pixels classified as forest only in the first
image à deforestation
2 for those pixels classified as forest only in the second
image à regeneration
3 forest in both the first and second images à
no change
|
