elements in total. The middle and southern 
arrays were in the same quarter scene. 
Locating ground element arrays in the 
images was optimized using regression. The 
approximate location of the arrays was known 
from fieldwork. Did the optimization procedure 
improve the correspondence between field and 
image data? Table 1 shows the differences in 
location for the three arrays, indicating a 
considerable decrease in residual variance by 
shifts of 0.5 to 3.5 pixels. 
Some prediction errors resulting from the 
evaluation are shown in Table 2. Three types 
of evaluation can be distinguished. 
The first type concerns the prediction of 
the training data. Results have been obtained 
using resubstitution, and, where applicable, 
using the lcave-one-out method. The latter 
results in higher estimates of the prediction 
error, but the differences are not large, and 
are not expected to be different for the other 
three methods. 
The second type concerns the prediction of 
pixels not included in the training set, but 
contained in the same quarter scene. This is 
the most relevant situation in practice. 
The third type concerns prediction of pixels 
in other quarter scenes. Strictly this is outside 
the scope of the present calibration procedure 
(radiometric corrections were not applied), but 
it is useful to compare the methods under 
very difficult conditions. Much higher values of 
RMSEP arc now observed. 
As could be expected, predictions of pixels 
of the same training set were most accurate, 
with RMSEP varying from 6 to 15%. Predicti 
ons of pixels from other training sets were less 
accurate, with RMSEP varying from 8 to 24%. 
Overall, a prediction error (RMSEP) of +/- 
15% seems feasible. 
The middle and southern arrays were 
added to form one training set, because they 
were both in the southern quarter scene. The 
northern array was used to calibrate the north 
ern quarter scene. The resulting regression 
equations for the northern and southern scene 
are listed in Tabel 3. 
Table 1. The effect of locational optimization. Shift and rotation of the pixel array, and decrease of 
residual variance relative to the unoptimized situation. 
shift (m) 
angle of 
rotation 
decrease in 
residual variance. 
north 
12.9 
0.8° 
-11.7% 
middle 
38.3 
3.0° 
-63.4% 
south 
86.0 
4.8° 
-77.4% 
Table 2: RMSEP in cover percentage. Where two numbers arc listed, the first is calculated by 
resubstilution, the second by the Leave-onc-out method. 
Training 
data 
Test 
data 
RMSEP 
heather 
RMSEP 
grass 
north 
north 
7.2/7.8 
6.3/6.9 
middle 
11.9 
10.7 
south 
24.0 
22.6 
middle 
north 
10.8 
7.6 
middle 
9.1/9.6 
8.2/8.6 
south 
17.2 
17.7 
south 
north 
26.2 
17.6 
middle 
12.0 
12.6 
south 
12.9/14.5 
13.2/14.9 
689