1018
can be derived from the training samples. The act
ual accuracies obtained after an analysis session
can be found using the test areas.
For a TM scene of 8000 by 8000 pixels and 7
channels, the processing times on LDIASl and LDIAS2
are too long. Therefore, we use the FMPS to reduce
the analysis time . Table 2 is a summary of the
FMPS performance for selected functions. These
functions are: 1) finding the histogram of all chan
nels of a full TM scene; 2) applying lookup tables;
3) computing ratios (normalization corresponds to
each channel being divided by the sum of all chan
nels); 4) parallelepiped classification, up to 256
classes; 5) maximum likelihood classification, up to
256 classes. These processing times are suffic
iently fast that the LDIAS project can achieve its
objective.
TABLE 2
FMPS PROCESSING PERFORMANCE
TM Image: 8000 lines by 8000 pixels
and 7 channels.
Function
Time
Histogram
Lookup Tables
Ratios (seven)
Parallelepiped classification
(256 classes)
Maximum Likelihood
Classification (32 classes)
10 minutes
20
30
35
150 minutes
3 A FORESTRY INTEGRATION EXPERIMENT
3.1 Methodology
The Canada Centre for Remote Sensing, with its CCRS
Image Analysis System (Goodenough 1979) developed in
the 70's, has demonstrated how remotely sensed data
can aid in the management of Canada's resources by
identifying geological structures, range-land
inventory, crop production estimation, forest cover
mapping, and environmental monitoring. CCRS, with
the support of the British Columbia Ministry of
Forests and Lands (BCMOFL), has conducted a series
of forest clear-cut recognition experiments to
demonstrate the feasibility of updating forest
inventory maps with imagery data. Mr. Frank Hegyi,
Director of the Planning and Inventory Branch,
BCMOLF, (Hegyi and Quinet 1983) supports these
experiments by providing the necessary digital
inventory maps, associated map labels, documenta
tion, and interpreted aerial photography. CCRS uses
this information to extract forest cover change
information from satellite, multispectral images,
and then updates the geographic information. The
updated information is returned to BCMOFL for eval
uation. The experiments have been so successful
that BCMOFL now uses LDIAS software and the proce
dures outlined below on its own computers.
The overview of the remote sensing analysis and
GIS integration approach is shown in Figure 3. The
GIS is assumed to be remote; that is, the GIS of
BCMOFL. The three main processing steps are: 1)
extract and process the GIS data; 2) analyze the
combination of the remote sensing imagery and geo
graphic information; 3) generate the updated
geographic information. The dashed line in Figure 3
shows the intended application area of our artifi
cial intelligence research. The elements of Figure
3 are expanded upon in Figures 4, 5, and 6.
The forest inventory digital maps used are at a
scale of 1:20,000 covering areas of 12.6 km by 12.6
km. The maps are in the UTM projection and are based
upon the topographic maps produced by British
Columbia's Survey and Mapping Branch. The attribute
information of these maps are stored separately in a
database at BCMOFL and are sent to CCRS as label
files. An overview of the geographic information
extraction and processing is shown in Figure 4.
Once these digital maps and label files are loaded
onto LDIAS2, the attribute information is extracted
and inserted into an attribute information database
(DMRS file) which is attached to the maps graphi
cally. The maps are displayed and the forest
classes of interest are identified. The maps
usually contain more than 256 classes. Therefore,
the forest classes must be grouped based on less
important attributes. This has not been a limita
tion yet, but it will be. For that reason, LDIAS is
being changed to handle a much higher number of
classes. The selected forest and non-forest classes
are then converted to raster or grid representation.
The gridded geographic information is used subse
quently for training and to mask out forested and
non-forested areas of the images.
The forest clear-cut detection is done with
single-date or multi-temporal images from the MSS or
the TM sensor, or both. Figure 5 is a summary of
the analysis methodology for multi-date, multi
sensor imagery. The images have been processed on
MOSAICS to eliminate radiometric errors and
rectified to the Universal Transverse Mercator (UTM)
projection. Rectification is performed by acquiring
ground control points from the image and calibrating
them to a UTM map. The map is produced by the
Surveys and Mapping Branch, Department of Energy,
Mines, and Resources. These maps are the standard
used for image rectification at CCRS. The map scale
is 1:50,000, and the image is rectified to a pixel
size of 25 meters. The gridded forest information
is combined with the remote sensing imagery.
Training areas can be selected manually based on
forest polygon classes or more recent ground infor
mation. The training or truth files are used to
generate the statistics required by subsequent clas
sifiers. The combined data set is classified within
the forested regions (based on the forest - non
forest mask). The classification is filtered using
a spatial, contextual filter (Goldberg and
Goodenough 1976). The classified image is assessed
for accuracy.
The GIS is updated following the procedure
outlined in Figure 6. The classified image is
converted to another grid format. This grid file is
changed to a vector representation with smooth
polygon boundaries and linkage to the DMRS database
of attribute information. The attributes are placed
in a label file, the graphics in a digital map file,
and the two files sent to BCMOFL for evaluation. In
the next two sections we report on one of our
experiments and the results obtained as an example
of GIS and IAS integration.
3.2 The Experiment
The particular imagery used for the experiment
reported here is a LANDSAT-5 TM image of an area
near Cranbrook, British Columbia in the Kootenay
Mountains. This cloud-free image was obtained on
August 15, 1984. The clear-cut areas lie on a river
valley at an altitude of about 1600m, surrounded by
mountain ridges up to 3000m. The forested areas are
covered mainly with spruce and balsam with ages
ranging from 150 to 200 years old. Non-coniferous
forest occupies less than 15 percent in any of the
forest polygons. The sizes of the clear-cut areas
vary from 12 he
the boundary p
18% to 43%.
The preprocess:
are:
A)
reflectiv
not used;
B)
normalize
banc
i
band i +
where i, j ar
Only four ba
were used as
correlated to g
Two methods
training method
training areas
The user judge<
geneous, perha
training area 1
ted training ar
by the user,
ing method wei
were 5 to 40 y<
less than 5 ye
greater than 4(
were performed
given previousl
The second
selection of p
except new cle
training methoi
than the data
Because of the
multi-modal cl
polygons for t
failed to pi
acceptable clas
3.3 Results
Over 1800 poly
map, but we ar
or less. Hen
database befoi
grouped into tl
of the follow
forest species
more of the pi
average age of
stratifications
productivity o]
the site condil
categories. Tb
ses, and then
in Section 3.2.
Table 3 list
cies obtained
preprocessing n
Table 3. Avera
Cursor Selecti
of Training Ar
Point Selectio
of GIS Polygon