location in the UTM co-ordinate system. Such counting "was done for no more meal 
than two or three segments of the fifty located. Precise placement is then totally 
done using the north-south and east-west survey pattern. segment 
| 
Training Data Selection and Classification: The selection of training data is ae 
a critical part of the project. In the Canola/rapeseed work, it was found in in sing 
several cases that training sets selected were inadequate and training had to segment 
be redone. The cases in which training data were inadequate invariably be anal 
resulted in classifications which missed whole fields of the crop or which both data in 
missed fields and included native vegetation. The signature for rapeseed is 
quite distinct in a region where the stage of growth is similar throughout | A 
(Brown, et al., 1980). In the Peace River District in 1981, however, growth | the: vi 
stage varied tremendously, in some cases from one farm field to another. | procedu 
Fortunately, even with this variability it was found that if a broad enough of: the 
training set. was established, most fields would be properly classified. The overlap 
problem was in defining an adequate training set. 
The method that evolved here employed the mosaiced 1:250 000 maps of 
segment locations to identify segments with Canola/rapeseed before any T 
analysis. The individual segment photo-maps annotated by field enumerators to 
n 
show field sizes and locations were then consulted to locate concentrations of GuconnE 
known Canola/rapeseed fields. These concentrations were in turn used to Missing 
identify potential subscenes for training. | reasons 
1981 a 
The training field selection during the subsequent analysis is typically Canada 
based upon subjective criteria. Since a number of fields showing a range of by clot 
Landsat intensity values is usually available in the potential scene one should missing 
at first select only those large fields which on the video display appear to of: data 
represent the range of intensities in fields stated to be rapeseed from the Ofssthi: 
training data. It should be noted that a knowledge of the crop phenology in segment 
the region and probable condition at the time of imaging is necessary at this 
stage to avoid including fields which may have been incorrectly identified in No Tra; 
the ground data, or which have changed since the field work was done. To avoid these r 
selecting atypical training data only a general location for potential training 
sites should be identified before one studies the video display data. I 
day, tl 
Once the training fields have been selected, the actual training set is where 
identified with the display cursor by keeping close to, but not including, the images 
field boundaries; bare areas and other crops should be excluded. The August 
parallelepiped classifier is then applied to the training set. The result is To. sol 
first checked against the fields used for training, and more training pixels or sites: 
fields can be added if warranted. If the result is "satisfactory", other fields 
fields not used in the original training are used to verify the classification. have ‘us 
The existence of a number of known fields in the same subscene as was used for approac 
training becomes important here. New training data can also be added from care, 
these fields if required. 
5 Isolate 
At this time, one must also ensure that there are no errors of commission thin he 
within the segments or, perhaps more importantly, in the non-agricultural thesen] 
lands, Where there are obvious errors of commission, several steps must be clear 
taken. First, if these errors consist of single pixels they may be ignored. complex 
Second, if it is but one crop, it should be noted in the reports for the obtain 
segment, and if confusion cannot be resolved, the result elsewhere should be classif 
checked to ensure consistency. If the errors are not consistent, the training of the 
set must be modified. If there are errors in the non-agricultural land, the vapour 
histogram limits should be adjusted to remove the confusion. If such Usually 
modifications affect the classification of known fields adversely, one can also The. sa 
associa 
subsect 
390 
NE S TRS