DESCRIPTION OF THE AUTOMATED PHOTO-INTERPRETATION SYSTEM

This experimental software allows for automatic tree species identification from digitized
aerial photographs. The images may be either black and white or color.  The unique feature
of this system is that stereo images may be used. It is recognized that stereo vision helps
considerably the photo-interpreter in identifying trees by allowing the perception of tree crown shape,
difficult to determine by using only 2D images.  This system uses digitized stereopairs to produce
a digital terrain model where every pair of common pixels are matched (occlusions are handled
by interpolation) by using a stereo matching algorithm in order to obtain a dense 3D
representation of the forest cover. A multipass approach is taken to permit crisp delineation of
shapes when computing heights.

Once this 3D model is produced, classification is obtained by computing texture, tone as well
as height "descriptive variables" from a moving window. The system works in a supervised
fashion: training sets must first be prepared by taking samples from already identified forest stands.
Then discriminant equations are computed. After, a confusion table is computed, allowing to check
the accuracy of the classification. Additional samples may be added as necessary. When the accuracy
is satisfactory, the user submit entire images to the system to be processed. Raster classification maps
are then obtained. The operation of this part of the system is very similar to a conventional satellite image
classification software except that texture and shape are used instead of only spectral signature.
Preliminary tests of the system shows that an accuracy of 95 % may be obtained for classifying
typical Quebec forest stands (boreal forest of Northeast North America).

Once the classification maps are obtained, they may be converted to vector format polygons
by using the included raster to vector conversion module. Optionally, very small polygons may
be regrouped to meet the minimal area requirements of forest inventory maps. This option can
be tailored by the user.

Version of February 1999
Jean Vezina
jvezina@attglobal.net (E-mail address updated 22 Aug 2000)