stanbul 2004 
getation. In 
s performed 
simulating a 
sification of 
nporal state 
, agronomic, 
sponding to 
the time in 
uired, t-/ a 
between t-/ 
ind £ is equal 
ellite image 
perform the 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004 
interpretation of the region of interest in the instant f, the 
proposed procedure employs: 1) Specific knowledge about 
the region of interest containing the classes of land use/cover 
and their respective spectral, contextual and multitemporal 
characterizations besides the rules that model the 
classification strategy; 2) Two registered images acquired by 
the same satellite in the instances f and t-/; 3) The accurate 
interpretation of the image in the instant t-/; 4) GIS data 
about the region of interest. 
The proposed interpretation procedure is based on segment 
classification. This allows the interpretation process to 
consider attributes such as shape, texture, and spatial relation 
between objects, in a more appropriate context for the 
representation of knowledge than the classic approach of 
pixel classification (Andrade, 2003; Darwish, 2003; Yan, 
2003). Therefore, the procedure begins with an image 
segmentation which outlines the regions with homogeneous 
spectral response in both input images. 
3.2 Specific knowledge 
In the context of the present research, specific knowledge 1s 
the knowledge background gathered both theoretically and 
empirically which tums an individual able, or more 
competent, to perform a specific task. In this case, the task is 
interpretation of low-resolution satellite images. 
Categories of knowledge 
In this work, the following categories of specific knowledge 
are required: 
1) Class description — Lists and describes the land use/cover 
(LULC) classes to be identified in the image. 
2) Spectral knowledge — Spectral signatures of the LULC 
classes 
3) Contextual Knowledge — Indicates the contextual 
information required for the discrimination of the LULC 
classes with similar spectral signatures. 
4) Multitemporal knowledge — Given the segment 
classification in #-1, relates its possible classifications in f and 
their respective possibilities. The possibility theory, 
introduced by Zadeh (1978), constitutes a context which 
allows treating the concepts of uncertainty in a non 
probabilistic way. 
5) Vector of attributes — vector composed by the 7 attributes 
considered by a photo-interpreter while classifying a 
segment. 
6) Rules of inference — Rules that describe the logic 
employed in the visual interpretation. Basically, it models the 
reasoning applied by the  photo-interpreter in the 
interpretation process by combining different items of 
spectral, contextual and multitemporal knowledge. 
The specific knowledge employed in the automatic 
interpretation of low-resolution satellite images can be 
acquired from experts on photo-interpretation, agronomy, 
ecology and, even, from the people of the region. 
The Automatic Extraction of Spectral Knowledge 
The contextual and multitemporal aspects of knowledge are, 
in general, dependent on the areas under analysis; however, 
they are independent of the image to be classified. On the 
other hand, the spectral knowledge is affected by the 
conditions in which the image was obtained. Climatic and 
atmospheric conditions, problems of sensor calibration, the 
level of soil humidity, etc., can lead to different spectral 
responses of the same LULC classes in images of the same 
regions, but acquired in different epochs. 
Usually, this difficulty is solved by supervised classification 
methods in which the spectral signatures are estimated 
considering samples selected by the photo-interpreter in the 
image to be classified. Nonetheless, this selection is usually 
manually performed. 
This section introduces a scheme designed to automate the 
procedure for estimation of spectral signatures (spectral 
knowledge). The dotted box in figure 1 encompasses the 
steps related to the automatic selection of training samples. 
Such procedure takes into account the input images, obtained 
in £ and t-1, and a reliable classification of the image of 7-1. 
First, the automatic selection of training samples submits the 
objects produced by the segmentation to an automatic change 
detection procedure which discriminates the stable and 
changed segments, considering the spectral responses in the 
instants £-7 and f. Then, the stable segments are used as a 
training set to estimate the spectral signatures of the different 
classes. 
The change detection approach assumes the following 
hypothesis: 1) The amount of segments changed between #-7 
and f is small; 2) The natural and man made events alters 
differently the spectral responses of the classes; 3) The 
images in t-7 and t are registered. 
3.3 Segments Classifier 
In this work, a fuzzy logic system is employed to model the 
reasoning of the photo-interpreter while performing the 
photo-interpretation, since this kind of system is a transparent 
model for modeling logical reasoning (Kuncheva, 2000). 
Mendel (1995) warns that, in general, in order to build fuzzy 
logic systems, the following is required: 1) employment of 
linguistic variables (Zadeh, 1965). 2) Quantification of the 
linguistic labels associated to the linguistic variables. 3) 
Definition of logical connectors, 4) the combination of the 
rules. 
In order to produce the interpretation of the scene in the 
instant f, the segments classifier employs: /) Specific 
knowledge; 2) low-resolution satellite images in £ and t-7; 3) 
the interpretation of the image in f-/; 4) the segments; 
5) maps of the region of interest.