Initially, it is necessary to define the linguistic labels, their
respective labels and the rules comprehended in the specific
knowledge. The modeling of the linguistic variables,
associated with contextual and multitemporal knowledge, is
performed manually. On the other hand, the fuzzy sets,
which define the spectral knowledge, are automatically
modeled.
Then, segment classification is performed. Initially, the
correspondent attributes vector is calculated. Then, the
inference machine, based on the inference rules, the fuzzy
sets, the rules and the attribute values, calculates the
membership values of the segment to each one of the classes
in the legend. The classification is given to the class with the
highest membership value.
4. EXPERIMENTAL RESULTS
The experiments presented in this section aim at evaluating
the potential of the proposed framework. The images
employed in the experiments are situated in the Taquari
Watershed, more exactly, in the County of Alcinópolis that
belongs to the State of Mato Grosso do Sul, Brazil. The
images were acquired on August 7, 2000; and August 10,
2001 by the satellite LANDSAT 7 (bands 3, 4 and 5).
The reference classification for this evaluation was produced
by visual interpretation by a photo interpreter experienced in
vegetal cover classification. In this procedure, it was
considered, besides the images of 2000 and 2001, the
classification of the image of 2000, the digital elevation
model, the drainage map and the photo-interpreter’s
knowledge about the region.
In the region of interest, the following LULC classes can be
found: Bare soil; Ancillary forest; Pasture; Water bodies;
Dense savannah; and Dense savannah in regeneration. For a
small segment of the input image, the figure 2 presents: a)
the supervised multispectral classification result; b) the
outcome produced by the proposed method; c) the reference
classification.
The results were assessed in terms of percentage of segments
that were correctly classified (classification ratio) and the
percentage of segments wrongly classified (classification
error) considering the previously mentioned reference
classification.
Table 1 presents the evaluation of the results produced by a
supervised multispectral classification and by the proposed
method for the entire image. The purpose of this experiment
is to evaluate the increment of the degree of automation
provided by the proposed approach in comparison to a
supervised multispectral classification.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
Supervised Knowledge
id based
Classification iin
Classification
Classification ratio 69 % 90 %
Classification error 31% 10 %
Table 1. Comparison of the results of the supervised
multispectral classification and the proposed methodology
The results shown in Table 1 indicate the superiority of the
results produced employing knowledge while compared to
the outcomes produced by the spectral classification. If both
results were given as basis to a photo-interpreter, he would
post-edit the classification of 31 96 of the segments
previously classified by supervised classifier, but only 10 %
of the segments classified by the knowledge based classifier.
This fact indicates that the use of knowledge can contribute
to incrementing productivity of the interpretation process.
5. CONCLUSION
This paper presented a framework for knowledge based
interpretation of multitemporal low-resolution satellite
images. The prominent points of the proposed methodology
are: its flexible structure which allows for straightforward
application of this model to low-resolution image
interpretation problems; and the automatic
learning/calibration of the spectral levels to the current time
image.
The proposed methodology was preliminarily evaluated
through experiments employing images of two regions inside
the watershed of the Taquari river, northeast of the State of
Mato Gosso do Sul. The evaluation showed that the
knowledge based results were superior to the spectral
classification results. This fact indicates that the use of
knowledge can contribute to the increment of the degree of
automation of interpretation process.
Multispectral classification using the manually selected
training set provided a classification ratio of 69 % for the
image of 2001. The remaining 31 % must be corrected during
post-editing. On the other hand, for the same image, the
proposed knowledge based classifier with the automatically
selected training set provided a classification ratio of 90 %.
In this case, only 10 % of the segments woul require post-
editing. Therefore, the quantity of segments whose
classification would be changed would decrease from 31 %
to 10 %. Additionally, considering that the proposed
procedure is capable of selecting automatically the training
set, the work of the photo-interpreter would be, in this case,
restricted to the post-edition of 10 % of the segments.
Ed Water bodies [] Unclassified
Figure 2. a) Supervised multispectral classification result; b) Proposed method outcome; c) reference classification
An
de |
e fc
Bra
pp-
Bik
GEC
An
Inte
Kin
Cle
Inte
Mu
Tra
779
Dai
Seg
In:
Syn
Kui
cla:
Cyt
Kui
for
In:
Bir
Lar
cla:
Coi
Lie
19€
Inte
the
anc
Ma
bas
Mc
Ba:
Pat
58:
Nic
ER
