Phase II: Satellite data analysis and process 
The adopted scheme for the process of satellite data follows classic 
criteria: the supervised classification and a maximum-likelihood classifier have 
been selected because of the possibility of a strict control on the activities in 
progress. The spectral and spatial resolution of the Thematic Mapper data at 
European latitudes, the particular accuracy in the training stage and In the 
preparation of the 'ground truth " and the availability of a sophisticated image 
analysis system (hardware plus software), lead to very good results. The 
processing scheme consists of six main steps: 
a) - Image selection and preliminary process. 
The selection of the satellite scenes on which to operate heavily depends 
on the availability of cloud-free images In the requested season. Previous studies 
show that, for forest mapping from Thematic Mapper data, the period from 
mid-June to mid-August offers a good compromise between appearance of the 
vegetation (phenology) and scene illuminance (due to sun azimuth and elevation). 
For classifying coniferous forest cover, winter data are much better than any 
other season, but the possibility of a spectral discrimination Is also good enough 
In summer, as shown in the accuracy evaluation section. 
Satellite scenes are acquired in the raw format stored on CCT. Different 
radiometric and geometric corrections are applied in the phases of the process. 
For classification purposes, the geometric correction doesn't include resampling 
with cubic convolution methods, which may affect the radiometric response. 
When necessary, athmosferic refraction and scattering effect (haze) can be 
locally removed with statistical algorithms. For the preliminary photographic 
output, oriented to the photointerpretation, more complex algorithms are used. 
b) - False colour image production 
False-colour images of the Landsat Thematic Mapper data for the entire 
area of interest are produced and output on photographic support for preliminary 
photointerpretation and area stratification. Among the possible Thematic Mappei 
band combinations the following have been selected (respectively in red, green, 
blue shadows): 
- 5-3-1 to enhance terrain morphology 
- 4-3-2 for vegetation monitoring 
- 7-4-3 for land-use. 
Geometric correction, with cubic convolution resampling and North-South 
image rotation, and edge enhancement is applied to these images to improve the 
readability and the allocation of the test areas. 
c) - Scene stratification 
In order to increase the accuracy of the classification, the image is divider 
in small zones of defined characteristics. This procedure is called 
"stratification" and the single zones "strata ". The aim is to divide the area intc 
subzones relatively homogeneous with respect to the spectral response; this 
should grant the extension of spectral signatures within the smaller area. The 
validity of a spectral signature is reduced to few kilometers when processing 
Thematic Mapper data over a montainous area. For each stratum a separate 
classification Is executed, using different signatures and different test areas. A 
post-classification analysis aggregates spectral classes that are 
stratum-dependent. Aerial photos, topographic maps and Landsat false colour 
Images (low and medium scale) are used as control data for breaking out these 
strata. Percentage of vegetation cover, percentage of bare soils, land features arc 
also used as stratification factors. The stratification is applied to the digital 
data stored on disk using a bit-map description language which masks the 
portions of image to be excluded from the process. This procedure increases the 
cost of processing phase because of the number of separate classifications and o' 
the final aggregation, but the stratification, when working over large areas, 
minimizes the "variance", or error due to sampling, in the final estimates. 
d) - Test area registration 
The portions of area for which "ground truth ' is available ( aerial photos ot 
tophographic maps interpretation, direct survays Information) are marked as 
"test areas". The allocation of such areas on the image is done manually on the 
available false colour prints, and interactively on the video display, where the 
image can be presented at the right scale and projection. 
e) - Stratum classification 
The classification stratum by stratum follows a classic supervised 
scheme, and consists of three passes: 
- selection of training sets over a test area, local evaluation and 
refinement; 
- test of the training set over other test areas of the stratum, evaluation 
and eventual iteration of first pass; 
- classification of the entire stratum and accuracy assessment. 
The procedure requires several iterations of the passes; a special 
emphasis has been therefore put, when designing and developing the Image 
analysis software package, in the contort of the man-machine interaction and ir 
the power of training set handling facilities. A maximum-likelihood classifier 
has been selected and used for the present application. The training areas were 
delineated on the display in a interactive manner. Decisions to merge or delete 
training sets were based upon the analysis of statistical parameters, of 
two-dimensional histograms and confusion matrices (see Tab. 3). A powerful 
software tools helped in individuating pixels In the training sets causing 
misclassifications and in their exclusion; this leads to spectral signatures thal 
’■locally" are as "pure" as possible. Several training sets for each class are 
extracted, trying to reproduce the various spectral aspects of a given category 
within the stratum. Density of the vegetation, sun exposition, terrain slope, haze 
presence are taken into account; they heavily affect the response of the 
vegetation. The result is a large number of spectral classes used for the 
classification and grouped later. The use of a powerful computer reduces the 
impact of such approach on process time. All the reflective Thematic Mapper 
bands (1 through 5 and 7) are used for the spectral analysis; the thermal channe 
(band 6) has been excluded because of the difficulty in the extraction of usefu 
information. 
f) - Class aggregation and final output 
The final aggregation of the classes across the strata, i.e. over the entire 
area will be performed using a computer look-up table procedure. The final image 
pixels, when classified, are assigne to one of the possible expected category. 
The classification tecnique is 'point-by-point", in which each pixel is 
treated individually. This approach produces maps which may contain even more 
detail than is actually needed. To avoid a "salt and pepper " effect in the present 
application isolate dots have been assigned to closer classes, according to a 
prevalence algorithm. 
The final classification is output on a film recorder, and printed at the 
scale of 1:100,000. Pixels assigned to the same class have the colour assigned tc 
the category bye the final legenda. Non-classified pixels are presented in 
Known 
category 
Number of 
pixels 
Percent 
correct 
Number of pixels 
assigned to category 
1 2 3 4 5 6 7 8 
1 
12 
91.6 
11 
0 
0 
1 
0 
0 
0 
0 
2 
22 
90.9 
1 
20 
0 
0 
0 
1 
0 
0 
3 
28 
96.4 
0 
0 
27 
1 
0 
0 
0 
0 
4 
11 
100 
0 
0 
0 
11 
0 
0 
0 
0 
5 
16 
87.5 
0 
0 
0 
0 
14 
0 
0 
2 
6 
20 
100 
0 
0 
0 
0 
0 
20 
0 
0 
7 
16 
100 
0 
0 
0 
0 
0 
0 
16 
0 
8 
22 
81.8 
1 
0 
0 
3 
0 
0 
0 
18 
TAB. 3 - Training sets confusion matrix over a test area 
1 conifers ,2-3 mixed, 4-7 conifers , 8 mixed