A neural network was trained to recognize two species and both sexes of mosquitoes. The frequency of the wingbeat is unique to each sex of each species. The neural network was given information about the wingbeat frequency and correctly classified the insects with a mean accuracy of 98%. Discriminant analysis had provided an accuracy rate of 84%. Even though the mosquitoes were of very similar species, the neural network had no trouble distinguishing them. Potential uses for this type of network include population/biological studies, pollination studies, evaluation of repellents and attractant, pest control, etc.
Aubrey Moore of the Maui Agricultural Research Station, University of Hawaii, developed this network to assess the feasibility of automatically identifying insects in flight. A photosensor was used to detect fluctuations in light intensity caused by reflections off individual mosquitoes flying through a light beam. Digital recording of the photosensor signals were made with an analog-to-digital recorder. A change in light intensity triggered storage of 512 samples. Each signal was converted to a 256-wide frequency spectrum using a Fast Fourier Transform. One input was assigned for each of the 256 spectrum slices. One output was defined for each of the sex/species combinations for a total of four outputs.
The training set used 403 samples, approximately 100 for each sex/species combination. The network was tested on 57 samples. The species and sex of every mosquito in the testing set was identified correctly by the network.
INPUTS
frequency content of wingbeat waveform:
0 to 19.5 Hz
19.5 to 39.1 Hz
...
(256 inputs)
...
4960.1 to 4980.5 Hz
4980.5 to 5000.0 Hz
OUTPUTS
male species A
female species A
male species B
female species B