M'Œdk 
reflectance in certain spectral bands, as recorded by 
a remote sensor channel (e.g., Landsat TM-4); 
altitude of a target (e.g., derived from a digital 
elevation model); 
ground truth data, such as temperature or humidity 
measured in situ-, 
derived attributes, such as the normalized vegetation 
index (Lillesand, 1987); 
requested attributes, such as the age of a forest or the 
degree of damage to a culture. 
Target attributes describe the general properties of this 
data and the paths or methods to access them. Target 
attributes are characterized by the following properties: 
- the role of the attribute with respect to an analysis, 
such as primary, intermediate, or final; 
- the possible values that may be assigned to the 
attribute; 
- the stability of the values of the attribute over time. 
(The elevation of a location may be considered stable 
over centuries, whereas other attributes such as 
temperatures may change within hours.) 
Target attributes are related to other knowledge-base 
entries, such as the target classes for which they are 
defined, or the processing models suited to compute the 
respective target attribute. 
Geographic Data 
The case-specific manifestations of target types and 
target attributes are represented by dynamically created 
objects, called geographic data. Geographic data are 
dependent on the geographic location and on the date of 
the analysis. The most prominent example of this kind of 
knowledge is the raster imagery recorded by remote 
sensors or derived by image-processing operations. But if 
these raster images were used as the only input into the 
data analysis, the results would be unsatisfactory. It has 
been shown that additional geographic information has to 
be taken into account as well (McKeown, 1987). There 
fore, other geographic data from the region to be analyzed 
must also be represented. Examples are: 
- geographic data describing the positions, shapes, and 
target classes of geographic units; 
- layers of a topographic map, converted into a compu 
ter readable format by a raster scanner; 
- a digital elevation model of the area of interest; 
- ground truth data; 
- results of former analyses. 
The geographic data relevant to remote sensing tasks in 
the environmental domain may be divided into two 
categories: classifications and attributions. Classifica 
tions and attributions may be considered as concrete 
manifestations of abstract target classes and target attri 
butes: 
- A classification may be characterized as a mapping 
(in the mathematical sense) defined on a part of the 
earth’s surface that associates locations with target 
classes. 
An attribution may be characterized as a mapping 
defined on a part of the earth’s surface that associates 
locations with values of a certain target attribute. 
Geographic data are represented by knowledge-base items 
that, in addition to a file of physical data, provide meta 
information for efficiently using the file. Geographic data 
include the following: 
- the data category (i.e., classification or attribution); 
- the related target classes or target attributes expres 
sed by the geographic data; 
- the name of a data file, usually containing the physi 
cal data in a raster-coded format as a matrix of data 
bytes. The raster format was chosen for most kinds of 
geographic data because it makes it easier to merge 
remote sensor data and ancillary data in the analysis; 
- encoding of the data bytes: a look-up table or formula 
that maps the interval between 0 and 255 onto the 
range of the data; 
- information about the coordinate system and the grid 
size of the raster data; 
- processing state: raw, histogram equalized, calibra 
ted, etc.; 
- a timestamp. 
Maps and images are similar to geographic data. They 
are stored in raster files as well, and may be computed 
using image-processing and spatial data handling proce 
dures. Unlike geographic data, however, they combine 
attributions, classifications and graphic elements in a 
form that is to be understood by a human rather than in 
terpreted by a computer. 
Processing Models 
Processing models are abstract descriptions of computa 
tions. They are static knowledge-base objects that do not 
change during a consultation. A model provides meta 
information for the use of traditional data processing 
operators, and it describes a set of input and output data 
and an algorithm. The algorithm is represented by a 
program which contains control structures and calls to the 
image processing and geographic database subsystems. 
Some of these models (such as the model for computing 
the vegetation index as a simple indicator for vegetation) 
are able to operate on pure remote sensor data. Other 
models (such as those for computing agricultural land- 
use classes by a supervised classification) need extra 
information (e.g., training areas), which may be provided 
by a geographic database or supplied by the user explicit 
ly. Some models describe the computation of geographic 
data to be used by a geographic information system, and 
other models describe how to visually enhance data to 
produce maps and images. 
149 
W'ï ,