ons of
those
'erall
tical
orted
ons.
Barrett, Rachel
3 RESULTS AND DISCUSSION
3.1 North West Coast 1997/98 Growing Season
Table 1 shows a range of predictive accuracy for the subject crops from 0% for triticale to 100% for oats. The average
predictive accuracy over all crops obtained from this assessment is 7096, with major crops such as potatoes, poppies,
pyrethrum, carrots, onions and peas approximately 76%. These data indicate that predictive accuracies should be greatly
improved by specifically tailoring interpretative images to cropping sequences.
Predicted
BA | OA | TR | WH| ON | €S | PE | BE | PO | PU | BR | CA | PP | PY
BA 5 0 0 1 0 0 0 0 0 1 0 0 0 0 7 71.4
OA 0 1 1 0 0 0 1 1 0 0 0 0 1 0 5 20.0]
TR 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0.0
WH | 0 0 1 4 0 0 0 0 0 0 0 0 0 0 5 80.0|
ON 1 0 0 0 16 | 0 0 1 0 0 0 0 0 0 18 88.9 E
CS 0 0 0 0 3 8 2 1 1 1 0 0 0 0 16 50.0} 3
. = PE 1 0 0 0 0 0 16 | Q 3 0 0 0 0 2 ]| 22 72.7 2
© | BE 0 0 0 0 0 0 0 3 1 0 0 0 0 0 4 75.0} 3
Po.j 301-001 "+3 43-12-0013" OT SO
PU 0 0 0 0 1 0 0 1 0 3 0 0 1 1 7 42.9} 2
BR 0 0 0 0 0 0 5 1 0 0 2 0 2 0 10 20.07
CA 0 0 0 0 0 0 0 0 0 0 0 4 0 1 S 80.0
PP 0 0 0 0 0 0 2 1 1 0 0 013412140 85.0}
PY 0 0 0 0 0 0 1 0 2 0 1 0 1 34 | 39 87.2
10 1 2 $21 9.1 30 {10.31 7 3 5. | 43 |-41. 153
50.0|100.0| 0.0] 80.0| 76.2] 88.9] 53.3| 30.0| 74.2] 42.9] 66.7) 80.0| 79.1| 82.9} 218 218
Producers Accuracy (%) 218 | 70.2%
Table 1. Omission/comission matrix for North West Coast 1997/98 growing season
Crops differ not only in phenology, but also in morphology and physiology, which affect the leaf area index and in turn,
reflectance (Nieuwenhuis et al, 1996). This point is particularly emphasised by the low accuracy with respect to the
identification of bean crops, which was 30%. Beans have a short growth cycle (approximately 8 to 10 weeks) but
conversely, a specific canopy closure period. As a result, only two images were of significant benefit in identification of
this crop, with the remaining 4 images, which were included, probably creating distortion within the programming
algorithms.
3.2 North West Coast 1998/99 Growing Season.
The methodology utilised in this project has resulted in some crops being identified with better than 90% accuracy
(Table 2). Favourable predictive outcomes have been achieved for crop types for which the greatest quantity of both
training and test data were acquired (poppies, 65.8%, pyrethrum. 86.4%, potatoes, 85.0%, peas, 78.0% and onions
69%). This reiterating that the choice of a good training set can also have significant influence on the success of a
classification approach, (Shine & Wakefield, 1999).
The overall accuracy for the 1998/99 growing season (69.8%) has declined from that recorded in the 1997/98 growing
season. Given that there was an increased number of observations made over the majority of crop types, this decline in
performance may be attributed to other factors.
The combination of too many variables and few observations (i.e. sample sizes) usually causes unreliable estimates
(Cruz-Castillo et al, 1994). It is therefore imperative that, if increases in accuracy of prediction are to be achieved,
sufficient calibration data (ground-truthing) must be obtained. There is however, a critical level where any increases in
the quantity of calibration data will no longer result in increases in accuracy.
Another possible contributing factor in the decline of the overall accuracy is the unsupervised classifier choice of
maximum likelihood. It is generally acknowledged that human photointerpreters use a considerable variety of
contextual information and common-sense experience in interpreting aerial imagery (Wilkinson & Burrill, 1991),
(Kontoes et al, 1993). Given that many classifier types are used in remote sensing (maximum likelihood, two-class
decision, parallelepiped, knowledge based, semi-variogram, fuzzy, kriging, etc.), much of this common-sense
experience comes into play with regard to selection of a classifier
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 137
