for 
1sed 
ics, 
Ore 
na 
1g, 
ve 
ed 
ic 
al 
n- 
as 
1e 
ge 
De 
interpreted as a knowledge based technique. In a strickter 
sense, however, information pertaining to the spatial and 
temporal characteristics, inter-connections and relationships 
of classes constitutes a real knowledge base. Such 
information allows the introduction of a-priori constraints 
and expectations and deductive reasoning into the analysis. 
Information which resides in a GIS can be incorporated 
either before classification for stratification, during 
classification for classification modification or after 
classification for post-classification editing. In stratification, 
cartographic data can be used to divide the scene into 
segments based on known administrative, property and land 
use boundaries, so that these sub-areas may be processed 
separately [Hutchison, 1982]. Polygon based spectral 
signature generation as samples for the classification is also 
possible. Classifier modification involves assigning a priori 
probabilities to the classifier, based on GIS stored 
information [Strahler, 1980]. Another approach is to 
perform per-polygon instead of per-pixel classification and 
use “IF-THEN” rule tables to reach a final decision based on 
a priori expectations and on the inputs from programs which 
model changes as a function of time. During 
postclassification sorting, GIS stored knowledge can help to 
resolve omissions and commissions [Hutchison, 1982]. 
3.2 Raster Image Statistics Generation 
CARIS/RIX has the capability of obtaining raster image 
statistics for existing polygons, supported by the vector 
polygon to raster object conversion software developed 
previously by Reedijk, [1990]. This software interactively 
creates raster representation of one or more map objects. 
The map objects may be explicit nodes, lines, or polygons. 
A region of a specified radius referenced to the map object 
may be included in the raster object (Le., this is a region 
specified by an "inclusion radius.") The map objects may be 
selected by pointing or windowing on the screen, or through 
a database query that refers to their lexical attributes. Map 
objects may also be defined by drawing a polygon with the 
cursor in the raster image backdrop. Boolean operations can 
be performed on raster objects. For example, a polygon and 
its perimeter can be separately rasterized, and then the latter 
subtracted from the formal to obtain a rasterized 
representation of the interior of the polygon that excludes a 
margin around its edge. Raster objects can be saved on disk 
in an object-based representation that is analogous to that 
used for vector-based data storage in CARIS. 
The fundamental problem of polygon to raster conversion is 
to determine which pixels lie within the bounds of the vector 
polygons. A simple but computationally expensive solution 
would involve the testing of each pixel to see if it lies within 
the polygon. A better approach is to use a polygon scan 
conversion technique. 
An extensively used algorithm to solve this problem is 
described by Foley and Van Dam [1984]. This method was 
slightly modified to resolve certain ambiguities. The basic 
algorithm is as follows: At each edge of the polygon, the 
points where it intersects the centres or each scan line are 
determined. This list of intersections is then sorted, first in 
y, then in x direction. Each pair of intersection points in this 
sorted list corresponds to the end points of a swath of pixels 
within the polygon. Several ambiguities in the Foley and 
Van Dam approach were resolved in the final 
implementation. This ultimately led to a solution where 
every pixel was assigned to one and only one polygon. 
Once the raster objects are defined and the image layers 
selected, the analyst can obtain the following image statistics 
for the corresponding map (vector) object: 
* mean vector and standard deviation, 
* variance-covariance matrix, 
* lower and upper value limits in each data layer, and 
977 
* the number of pixels in each map object or groups of map 
objects. The statistics can be generated for single map 
objects or for a group of objects belonging to the same class. 
All image statistics are stored in the database as attributes of 
the map object. The frequency histogram of each layer can 
also be displayed. 
The training statistics support the minimum distance to 
means, parallelepiped, and Gaussian maximum likelihood 
classification decision rules. A trial classification and a full 
classification option are available. 
The trial classification option operates only on the raster 
objects selected as sites for the training statistics generation. 
The output is a confusion matrix and a theme assignment for 
each class, as well as for the unclassified and multiclassified 
pixels. The full classification option operates on the entire 
data file stored on disk. It creates theme files and pixel 
counts for each theme as the output. 
Image statistics generation within pre-defined or user defined 
polygons has distinct advantages over the customary training 
statistics generation performed in image analysis systems. 
The statistics obtained can be stored as an attribute of 
individual polygons. It can aid qualitative decision making 
and can facilitate change detection without performing the 
classification of multitemporal data sets. Polygons càn be 
located by codes, whereby potential difficulties and errors in 
the identification of training areas are eliminated. Operating 
within a GIS provides the means for developing per-polygon 
and contextual classification schemes and for including 
textual attributes in the analysis.  Pre-classification 
segmentation and masking can be based on administrative or 
property boundaries and land use information residing in the 
map file. The combined vector-raster approach also 
facilitates the utilization of image data in corridor analysis. 
3.3 Example 
The knowledge based image analysis scheme is illustrated by 
a simple agricultural crop classification project. The 
Department of Agriculture and various marketing boards 
need annual forecasts on the yield of crop production and 
also wish to monitor changes in agricultural land use. 
Spaceborne images, such as LANDSAT-TM or SPOT data, 
are an excellent source of information for this purpose. A 
typical procedure for the execution of this project, in a stand 
allows image analysis, is as follows: 
Near the time of the satellite's overpass, several fields are 
selected for each crop to be inventoried to serve as training 
fields during the analysis. The selection is made by site 
visits to assure total reliability. These fields are marked on 
maps or aerial photographs. After the satellite data has been 
acquired, these fields are visually identified in the image 
display on the screen of the analysis system and manually 
delineated with the cursor and stored in a training area file. 
The image statistics of each crop is then generated and used 
as an input to a per-pixel supervised classifier. Post 
classification edit is then performed to remove "stray" pixels 
and fill small unclassified gaps within the fields. Finally the 
total area covered by each crop in the region is determined by 
pixel count. It is also common to improve the area estimate 
by regression analysis. Permanent visual records may be 
produced in a suitable dot matrix plotter. 
The knowledge based analysis approach to the same project 
would be implemented as follows: It is presupposed that a 
digital map base is available covering the region of interest. 
Administrative, property and farm filed boundaries are 
acquired in digital form and added to the digital map file in 
the GIS. An identification code is assigned to each field in 
the polygon attribute file. Training samples are collected by 
site visits in the usual manner but are stored by field 
identification codes in the data base. A list of crops