Mar gs whit Tr 
FERSHED A full description of the treatment of Landsat data in this project is given in 
/5/. Below a summary follows: 
a TO 
vedish 
logical 
1S 
ympared 
To fully utilize available information for obtaining watershed data, map data 
has been included in a maximum-Tikelihood classification of the Landsat data. 
Overlays of maps in 1:250 000 have been digitized with an Optronics scanner, 
transformed to the map grid system and converted to 50x50 meter pixels. Those 
overlays which have been used are masks representing water, forests, bogs 
and built-up areas. 
Also Landsat-data has been geometrically transformed to this coordinate system 
and the data has been resampled to the same pixel size. 
Two Landsat-scenes from adjacent paths covered the watershed recorded two 
following days in the end of May 1978. Training statistics for classification 
was obtained by using aerial photography, maps and a graphical image display. 
) For arable land clustering methods were used to obtain the statistics. 
The Landsat-data has been categorized into 34 classes. The map data was used 
in the following way: To every class on the map are connected a number of 
classes representing training areas (candidate classes in the classification). 
In the computations the classifier is first directed to classify among those 
candidate classes indicated by the map class. The class with the largest 
probability is tested if it is within given confidence levels. If so, the 
choosen class is accepted - otherwise the classifier investigates all candidate 
classes for the most probable class and accepts that class if it is within 
the confidence levels. The classifier may be directed to pronounce the infor- 
1 od mation from the map if the analyst chooses a large confidence interval for 
the acceptance and vice versa. 
  
  
ACCURACY 
The agreement between map classes and classes obtained from the classification 
was investigated for 2.9 milj. pixels. This showed that 83 percent of the 
pixels were in aggreement with the map. The following table compares Landsat 
classification with digitized maps in the scale 1:250 000. 
CLASSES AREA PIXELWISE 
(percent) AGGREEMENT 
(percent) 
CLASSIFICATION MAP DATA 
3 WATER 12 12 92 
] BUILT-UP AREAS 4 2 71 
FORESTS 51 45 94 
BOGS 4 4 69 
ARABLE LAND 27 37 69 
REJECTS 1 
In the maps room has been left for major roads and these areas will appear 
as arable land in the digital data. Thus arable land in the map data should 
be reduced and forest areas increased to some extent. 
In order to verify the results of Landsat-data classification, the data has 
also been compared to 'ground truth' land use data obtained from digitized 
— topographic maps in the scale 1:50 000. The following land use classes have 
been studied: water, forest and bogs (those which were obtainable from the 
digitized maps). Cluster analysis have been applied to the series of areal 
subcatchment values of the named parameters, obtained by conventional and 
  
411 
P s ; 
^. ee nn EEE QE EE EE lll eia at ilius