Bio-Applied Systems Engineering

Contract PH0191 Generic System for Detection of Statutory Pests and Pathogens

Introduction

As part of ongoing research into the recognition and identification of quarantined insect pests (DEFRA Project PH0191), recording techniques and identification methods have been established to discriminate between the biting sounds in timber of several insect species.

One of the major insect pest families to be investigated in this research project was the Cerambycidae, with a particular concern in recognising Anopholora glabripennis (Asian longhorn beetle). Due to the availability and licensing requirements of this species initial studies were carried out using another quarantine species and two indigenous species; these being Hylotrupes bajulus, Prionus coriarius, and Rhagium bifasciatum respectively. The techniques and results presented here are with respect to these species.

Detection

A great deal of research has been carried out in insect acoustics to establish its suitability as a diagnostic tool for use in such areas as biodiversity assessment, behavioural studies, and environmental monitoring to name but a few.

However, much of this recording work has a tendency to be laboratory based and where work is extended to field applications these are ordinarily limited to short term studies. This is in part due to the nature of the equipment used to carry out such recordings. Firstly, recording equipment used is generally expensive, can be bulky, and is susceptible to environmental conditions, thus restricting where it can practically be used. Secondly, current technology limits the amount of data that can be collected.

In addition, the majority of research concentrates on non-incidental acoustic airborne signals, such as stridulation, when used as behavioural responses. This obviously limits the usefulness of such recordings for detecting pest species since this method relies on insect interactions. The approach adopted here is to investigate incidental acoustic signals from such activities as moving or feeding, which should be more abundant.

A range of sensors have been investigated to determine their suitability for recording incidental sounds while addressing the issues of cost, portability and robustness. From these investigations three sensor types have been identified: electret microphones, bi-morphs, and piezoelectric disks. These fall into two categories: acoustic and vibration sensors.

Summary

The approach adopted for this project was to develop hardware solutions to overcome the problems outlined above. This has been achieved with respect to the actual recording devices; with further research to improve the storage capabilities (see future work).

• Utilisation of low cost sensors: electret microphones (£2.50), bi-morphs (£1.76), and piezoelectric disks (£0.35).
• Significant cost reduction in comparison to standard audio and vibration sensors; average cost of microphone £140, and accelerometers £350.
• Amplification hardware built into a hand held unit for use with digital audio tape (DAT) and mini disk recorders.
• Current limitations on data acquisition are dependant on tape and disk capacities.

Typical Bite Waveforms For Hylotrupes bajulus, Prionus coriarius, and Rhagium bifasciatum

Recognition

After the collection of sound or vibration data has been completed the waveforms obtained require analysis to determine the identity of the target species. This is a two stage process that involves encoding the waveform signal and then presenting the encoded data to an artificial neural network.

Data analysis is carried out using Time Domain Signal Coding (TDSC) which is a technique that analyses specific temporal characteristics to encode a signal; these are the duration and the shape of waveform segments called epochs. The duration is the time interval between consecutive zero crossing points and the shape is determined by the number of positive minima or negative maxima between the zero crossing points.

Time Domain Signal Coding (TDSC) Encoding Techniques

The duration (D) and shape (S) of each epoch provide a D & S code pair. The D-S code pairs are then mapped non-linearly onto a fixed size code book which reduces the number of representative symbols produce by the encoding scheme.

Once the wave form has been encoded using TDSC the encoded data can be analysed as either a frequency distribution of codes, or as a frequency distribution of pairs of codes, to produce either an S-matrix or an A-matrix respectively. The S-matrix is a one dimensional histogram based on the number of individual D-S code combinations whereas the A-matrix is a two dimensional representation of pairs of codes i.e. how often code ‘X’ follows code ‘Y’.

Typical A-Matrices for Hylotrupes bajulus, Prionus coriarius, and Rhagium bifasciatum

Recognition of the insects is achieved by presenting the S or A Matrices, which are unique for each sound, to an artificial neural network (ANN). This ANN can be trained to recognise certain patterns from the matrices and compare them with known data to give a result.

Results

Recordings were made of the three species using piezoelectric disks as sensors. The recordings were then analysed and number of example of bites for each insect were extracted from the waveform signal; these examples were used to generate a training data set and test data set. The training data sets were used to train a Multilayer Perceptron network; the test data sets were presented to the network and the following classification table obtained.

The results show all bites from R. bifasciatum were correctly identified, with 86% of bites from H. bajulus and 80% of bites from P. coriarius were correctly identified. The results also show that all the bites from H. bajulus and P. coriarius that were misclassified were confused with R. bifasciatum. However, these results are only preliminary and from a limited data set.

Future Work

This poster has outlined the progress made in the development of both the hardware used to obtain insect recordings and the software used to recognise which species have been recorded. This work can be further improved with the following considerations.

• Data Collection: Development of data collection method that will enable long term recording to be carried out. Currently around 90% of recorded data contains redundant information. This problem is being tackled with the development of systems that will only record data of interest and systems which can perform real time analysis of data and record pertinent information.

• Acquisition of larger data sets: Currently recorded data is limited to a small number of insects and host materials. To further develop the methods outlined requires greater number of insect recordings to be obtained. This is being achieved through collaboration with CSL, PHSI and Royal Holloway

• This research was presentad as a poster at the Royal Entomological Society Symposium "Insect Conservation Biology and Annual National Meeting" University of Sussex, 12-14th September 2005
  
  
  
  (Download as pdf 2Mb)