8.5. INTEGRATED PEST MANAGEMENT SYSTEMS 281
ïIf ratio of insects is small and pest A is high, then apply high concentration
pesticide Z.ïIf ratio of insects is medium and pest A is medium, then apply medium
concentration pesticide Z.While this example does not specifically discuss the formulation of such
decision-making rules, we are merely suggesting that it is conceivable that such
rules can be generated. The significant attributes of such decision rules is based
upon the accuracy of insect classification and how well we can actually determine
the pest populations in terms of specific species.
Note also that the objective is to perform such classification in an automated
manner with little or no human intervention. The intent is to developfield de-
ployable IPM systems where there is little or no access to experts in entomology.
As such, the IPM system is expected to provide the expertise to the farmer di-
rectly in an agricultural setting. Classical statistical methods fail to provide the
necessary basis for such an approach to developing successful IPM systems.
There are several important functions that are desired in a vision-based clas-
sification system. First, an imaging systemis needed to obtain high resolution
images of the objects requiring classification. This can be obtained using a high
resolution digital camera.
The second function is an image processing system that is capable of seg-
menting the original image, performing thresholding to eliminate debris and
other unwanted material in the sample collection, and providing images of in-
sects at a high enough resolution to allow appropriate feature extraction.
The third function is the classification system. Statistical methods suffer
from their inability to provide a model that can account for the large variations
that can occur in the insect morphology. Nontraditional approaches such as
artificial neural networks, fuzzy logic, and possibly other hybrid approaches
need to be investigated for classifier design.
Classifier design Pattern classification is generally considered to be a high-
end task for computer-based image analysis. The complexities range from locat-
ing and recognizing isolated known objects, recognizing objects in poorly defined
classes, to much more open-ended problems of recognizing possible overlapping
objects or classes. The idea therefore is to select an appropriate set of features
that is common to all object classes and that can provide discrimination be-
tween various classes. Ideally, one would expect the features to be distinct for
each class in order to obtain ìperfectî classification. Obviously, in a less than
ideal environment such ìperfectî classification cannot be obtained.
Table 8.18 provides a list of insects that inhabit cotton and alfalfa agricul-
tural ecosystems, and that are referred to by their common names.