Real-time acoustic identification of invasive wood-boring beetles

Wood-boring beetles are a cause of significant economic and environmental cost across the world. A number of species which are not currently found in the United Kingdom are constantly at risk of being accidentally imported due to the volume of global trade in trees and timber. The species which are of particular concern are the Asian Longhorn (Anoplophora glabripennis), Citrus Longhorn (A. chinensis) and Emerald Ash Borer (Agrilus planipennis). The Food and Environment Research Agency’s plant health inspectors currently manually inspect high risk material at the point of import. The development of methods which will enable them to increase the probability of detection of infestation in imported material are therefore highly sought after. This thesis describes research into improving acoustic larvae detection and species identification methods, and the development of a real-time system incorporating them.

The detection algorithm is based upon fractal dimension analysis and has been shown to outperform previously used short-time energy based detection. This is the first time
such a detection method has been applied to the analysis of insect sourced sounds. The species identification method combines a time domain feature extraction technique based
upon the relational tree representation of discrete waveforms and classification using artificial neural networks. Classification between two species, A. glabripennis and H. bajulus, can be performed with 92% accuracy using Multilayer Perceptron and 96.5% accuracy using Linear Vector Quantisation networks. Classification between three species can be performed with 88.8% accuracy using LVQ.

A real-time hand-held PC based system incorporating these methods has been developed and supplied to FERA for further testing. This system uses a combination of dual piezo-electric based USB connected sensors and custom written software which can be used to analyse live recordings of larvae in real-time or use previously recorded data.