[IC 
[mage 
tions, 
image 
ent of 
shold 
assed 
those 
es are 
lue to 
‘or an 
image 
abase 
IS) is 
raster 
lyzed 
. The 
) used 
renyi 
pixel 
ecific 
BMS, 
ts can 
sented 
toring 
ch are 
Final 
ed on 
on the 
layers 
g the 
ent of 
n. 
ormed 
3 done 
parate 
al, but 
laps is 
clearly 
separated in the files. An advantage of operating in a GIS is 
that polygon attributes are good sources as reference data 
and the training samples can easily be located by the 
polygon identification numbers. The processing of the data 
one theme at a time allows the use of the parallelepiped 
classifier, which is the simplest algorithm. 
Upon completion of the pre-pixel supervised 
classification, the multi-layered output is analyzed polygon- 
by-polygon, one layer at a time. The polygons are selected 
for assessment either interactively by pointing on them 
with the cursor on the screen, or in batch processing mode 
by selecting them through the polygon identification codes. 
A pixel count is then initiated in each polygon to determine 
the number of the pixels in a particular polygon that are 
labeled as belonging to the various classes. At this point a 
decision has to be made on the final class assignment of the 
polygon being assessed, assisted by the discipline specific 
expertise of the analyst, and on ancillary information 
available in the data base. 
Various alternatives could now arise. For example, if the 
percentage of pixels belonging to a particular class is above 
a preset threshold, then the entire polygon area is declared 
as belonging to this class. This means that this area 
remained unchanged or the change occurred according to 
expectation. If the polygon contains pixels of more than 
one class in significant percentages, then it is declared as a 
dual or multiclass area. The proportion of the class 
distribution can be adjusted to the known area of the 
polygon as a constraint. Polygons with multiclass pixels, 
none of which add up to a minimum threshold level, or with 
a significant number of unclassified pixels, are flagged for 
further examination. Further data base query may resolve the 
issue or on-site inspection may be necessary. 
In the second analysis method of the raster data, the per- 
pixel image classification is by-passed entirely, and only 
image statistics are generated polygon-by-polygon. Mean, 
standard deviation, median and mode are likely candidates. 
A comparison of these statistics, one theme at a time, with 
those obtained for the training samples, forms the basis of 
the class assignment of each polygon. Various alternatives 
can again emerge and the final decision requires project 
specific expertise. This method is a one step operation. 
3. IMPLEMENTATION 
The polygon specific classification scheme was tested on 
monitoring the growth of forest plantations using Landsat 
Thematic Mapper (TM) imagery. The 367 km? study area 
contained plantations of two coniferous species labeled as 
PL1 and PL2. There were a total of 225 plantation plots 
ranging in size from 400 ha to 16,000 ha. Planting was 
done between 1977 and 1993. The objective was to check 
whether the development of the individual plantation stands 
followed the expectation. Stands that diverged from the 
normal development curve beyond a certain tolerance were 
flagged for further examination. The biophysical basis of 
this monitoring is the assumption that the percentage of 
crown closure of the trees is a valid indicator of the 
development of the plantation [Honer, 1972; Danson, 
1984], and that the spectral reflectance received by the TM 
Sensor is a function of the crown closure [Gemmell, 1995; 
Leckie et al., 1992; Spanner et al, 1990; Stenback and 
Congalton, 1990]. The study area is located in north-central 
New Brunswick. 
The Landsat TM image was recorded in November 1994. It 
was cloud free and of good quality. The boundaries of the 
plantation stands were shown on a 1:50,000 scale 
plantation map, and were manually digitized. A database 
was then set up in the GIS, where each plantation polygon 
was assigned an identification number and the year of 
planting was recorded as an attribute. 
Geometric registration was handled in a novel way. The 
objective of this project was a polygon specific monitoring 
of plantation development. Thus it was sufficient to have an 
accurate geometric registration of the raster image and the 
vector polygon layers in relative sense, and registration to a 
map grid in an absolute sense, with the help of ground 
control points, was not necessary. Therefore, the polygon 
network was transformed to fit the image. This was a much 
faster and simpler operation then the geometric and 
radiometric transformation of the whole multiband image 
file, and was easily accomplished in the GIS [Derenyi, 
1994]. 
Training samples were selected with the help of the 
polygon overlay. The TM image was displayed on the 
screen with the plantation boundaries superimposed, and 
groups of representative pixels were delineated for each 
theme within designated polygons. The polygons selected 
for training in each plantation age group were from among 
those where the normal expected tree growth was confirmed. 
The PL1 plantation species consisted of seven age groups, 
planted in the years: 1977, 1978, 1979, 1981, 1987, 1990, 
and 1993. The PL2 species were planted in the following 
six: years: :1977,-1978,: 1979,:1980, 1981, .and 1992. The 
image statistics generated for each age class were: the 
minimum and maximum digital numbers (DN), the mean 
value, standard deviation, median and the mode. TM spectral 
bands 3, 4 and 5 were only used in the analysis. 
The next step in the per-pixel image classification 
method was to create the thirteen theme layers by 
parallelepiped classification, performed one theme at a time. 
The decision region was set at three standard deviation. 
"Thereafter, the percentages of correctly classified pixels 
were computed in each polygon. Correctly classified meant 
that the pixel in question had been placed in the growth 
class which corresponded to the plantation year of that 
polygon. In other words, the classification of this pixel 
agreed with the apriori expectation. Such was the case for 
example, if a pixel, located in the polygon with attribute 
PL1-77 (species PL1 planted in 1977), also resided in the 
PL1-77 information class (theme) layer. 
The mechanics of the assessment was handled as follows: 
All polygons belonging to the same plantation year were 
retrieved one-by-one through a database query. Each vector 
polygon was converted into a raster object and overlaid on 
the corresponding information class layer obtained through 
the image classification. The degree of agreement of the two 
layers was measured by a pixel a count 
The results of the assessment were grouped into four 
categories according to the following percentage ranges of 
agreement with the apriori expectation: 
0-30%, labeled Very Low (VL), signaled major problems. 
The majority of the threes were probably lost, severely 
retarded in growth or perhaps there was an error in the 
database entry. 
31-50%, labeled Low (L), indicated unsatisfactory 
development of the plantation. The stand may be under 
213 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996