1016 
to substantial improvements in performance. 
A second approach would be to use an existing GIS, 
usually in vector form, with raster image data and 
classes as specified by the resource manager. In 
Canada, resource management agencies have, over many 
years, made large expenditures building geographic 
information systems. In British Columbia, for 
example, 3000 forest cover maps, out of 6000 
possible, are in digital form in a forest inventory 
GIS. Agencies with large data bases are unwilling 
to change to a new GIS representation in order to 
make use of remote sensing data. Therefore, remote 
sensing image analysis systems in Canada must inter 
face to existing GIS. This and other differences 
lead to several problems given in the next section. 
One way of exchanging geographic information 
amongst several GIS is to use an accepted standard 
for information exchange. In several countries 
there are attempts to define and implement GIS 
exchange formats. Suppliers of GIS have also speci 
fied exchange formats. In Canada, an early attempt 
was made by several federal departments to produce a 
GIS exchange format, the Spatial Data Transfer 
Format (Goodenough et al, 1983). This format is a 
member of the LANDSAT Ground Station Operators' 
Working Group (LGSOWG) family of tape formats for 
remote sensing imagery. Mapping agencies in Canada 
are cooperating to develop a more extensive format 
for topographic data exchange. Even between GIS 
from the same supplier, there are labelling incom 
patibilities between different GIS. There needs to 
be, but currently does not exist, an international 
standard for geographic information exchange. 
1.3 Problems Integrating IAS and GIS 
Most GIS use vector storage to represent thematic 
classes. Image analysis systems use raster storage. 
An integrated GIS-IAS must have software to convert 
from raster to vector, and vice versa. The grid 
resolution of the remote sensing image may not allow 
one to reproduce the fine structure of the GIS. 
A GIS can have thousands of classes. For example, 
a forest cover map can have more than 1800 different 
classes. Most IAS handle less than 256 classes 
since the most popular classification algorithms 
operate with a cost proportional to n 2 where n is 
the number of classes. The LDIAS analysis subsystem 
is restricted to 256 classes. For forestry analysis 
on LDIAS it is necessary to combine classes based on 
ranges of attributes, such as slope, stand density, 
species, or site quality. 
The geometric accuracy of the GIS is probably 
worse than current, corrected remote sensing data. 
We have found that geocoded Thematic Mapper data 
with an IFOV (instantaneous Field of View) of 30m 
and a resampled pixel size of 25m are more precise 
geometrically than several geographic information 
systems. This is, in part, a result of the the much 
greater synoptic coverage of a LANDSAT scene (185 km 
by 185 km). The GIS are usually derived from maps 
or combinations of maps and aerial photography. The 
maps have been derived using aerial photography. It 
is not uncommon for stereo models derived from the 
aerial photography to have displacement errors in 
rugged terrain where there are few man-made 
features. 
Another problem is that the GIS class labels may 
not correspond to detectable remote sensing classes. 
For some classes, the satellite data we are using 
does not have sufficient resolution to detect, for 
example, small creeks overgrown with trees or insect 
damage for individual trees. Some classes require 
the use of contextual knowledge; for example, parks, 
recreation areas, cemeteries, etc. Some classes 
easily identified with a remote sensing IAS may not 
correspond to classes desired by the resource 
manager, such as a combination of several forest 
species which form a single spectral group. 
Even if the GIS and IAS are in grid or raster form 
for analysis, the methods of interpretation to 
choose are dependent upon the sets of classes being 
examined. A few spectral channels may suffice to 
identify major water bodies, but many spectral and 
spatial features may be needed to classify some 
forest species. Optimization is important here 
since the processing and storage costs can rise 
dramatically with the number of channels used in the 
analysis. 
For operational resource management, the GIS and 
IAS are complex, large, computer systems. Both 
systems require large storage capacities, many 
software functions, and complex displays. Analysts 
who use such systems have received at least one year 
of training. There are major problems relating maps 
to images, as given in sections 1.3 and 4. These 
problems require symbolic reasoning (artificial 
intelligence) for their solution. The use of expert 
systems simplifies the user interface, but adds to 
the system complexity. Expert system applications 
are discussed in section 5. 
2 THE LANDSAT DIGITAL IMAGE ANALYSIS SYSTEM (LDIAS) 
2.1 The LDIAS Objective 
When the United States of America launched the 
LANDSAT-4 remote sensing satellite, the Canada 
Centre for Remote Sensing began receiving data from 
the two sensors, the Thematic Mapper (TM) and the 
Multispectral Scanner (MSS). The TM sensor produces 
images with 10 times more data than the MSS. The 
CCRS Image Analysis System (Goodenough, 1977), 
designed for MSS analysis, could not handle a full 
TM scene. Larger processing and storage power were 
required, as well as new functions for analysis with 
higher spatial resolution imagery. 
In 1982, CCRS received approval to begin a 
research and development project for TM analysis. 
The objective was to conduct this research and 
create a new image analysis system, the LANDSAT 
Digital Image Analysis System (LDIAS) , which would 
enable us to analyze a full TM scene (185 km by 185 
km) into 32 classes within 8 hours, while permitting 
the integration of map-based data. In addition, 
LDIAS would have to support airborne optical scan 
ners and synthetic aperture radars. From this 
general objective there developed many activities 
which led to the system architecture and functional 
ity described below. 
2.2 System Architecture 
The architecture of LDIAS is shown in Figure 2. The 
primary image analysis computer, LDIASl, is a Digit 
al Equipment Corporation (DEC) VAX 11/785 with 8 
megabytes of memory and 2.6 gigabytes of disk stora 
ge. There are four image displays, each with 512 by 
512 pixels displayed: two Gould Deanza IP8500 dis 
plays (12 channels), and two Dipix Aries II displays 
(4 channels). All terminals connect to a Gandalf 
Compacx switching network which permits one to work 
on any computer from the same terminal. The next 
largest computer in LDIAS is LDIAS2, a DEC VAX 
11/780 with 8 megabytes of memory and 1.6 gigabytes 
of disk storage. This Map Input/Output subsystem 
was acquired from Intergraph Corporation. There are 
two color Interact map displays, each with two 
screens and digitizing tablets. To speed graphics