In-orchard imaging of pecan weevil and efficient identification using orthogonal polar moment descriptors

Abstract

The pecan weevil is considered the most harmful late season pest of pecan and requires population monitoring in orchards for proper pest management practices. In this research, algorithms for detecting and identifying pecan weevil are developed based on machine vision and pattern recognition components. Instrumented traps were designed and constructed to collect images of insects as they enter a pecan weevil trap. The instrumented traps were fitted to trees in a pecan orchard and left to collect images through a late pecan growing season. These images were processed by computer algorithm to be uniform, then used to test and train an insect detection classifier system that can predict whether an insect is present in the trap for each image.

Images containing insects were further processed to extract only the shape of insect silhouettes. This silhouettes were then used to extract Zernike Moment, Pseudo-Zernike Moment, Fourier-Mellin Moment, and MPEG Angular Rotary Transformation shape description features. The shape descriptors were combined into feature vectors before being used to train classifiers that can discern pecan weevil from non-pecan weevil insects. Pecan weevil identification was shown to be over 98% accuracy depending on the shape descriptors used.

Two algorithms for reducing the number of features in the shape descriptor feature vector were developed based on concepts of Principal Component Analysis and Fisher Multiple Discriminant Analysis. The two algorithms were then used to select the features from the entire set of insect shape descriptors in order to reduce the number of features that need to be stored and used in classification. These two methods were able to reduce the number of shape descriptors from over 1000 to as few as 25 while maintaining classification accuracy above 95%.