The Insect Sound Classification project investigates the creation of a similarity measure and recognition system for insect flight sounds. We have implemented a mobile system consisting of an Android app and a server classification component. The technology is designed to be used to identify insects by their sounds in situations where taking images is difficult or impossible (e.g., at night).
We participated in an insect classification contest held by Dr. Eamonn Keogh's lab at the University of California at Riverside (UCR) to create a similarity measure for insect flight sounds. According to Dr. Keogh “the task is inspired by two socially noble problems. If insects could be correctly identified using cheap sensors, one could: (1) plan malaria interventions more effectively, perhaps saving some of the million human lives lost each year, and (2) plan insect pest crop interventions more effectively, thus growing more food with less pollution, less energy, less environmental damage, etc.” These goals match excellently with the COSMIC centre’s vision and we have continued on this work by implementing an insect sound classification system consisting of a mobile Android app and a server classification component.
- Identify insects through their sounds in order to carry out, e.g., malaria interventions to help save human lives
- Effectively conduct crop pest interventions to help grow more food with less pesticides, less energy and less environmental damage
- Some insects directly generate sounds
- Different insects have different flying patterns, which leads to air vibration and indirect sounds
- Typical spectral features (STFT, Mel-spectrum and MFCC) are concatenated with proper weights
- SVMhmm is used to project the feature sequence into a new vector
- Each dimension of the new vector corresponds to one class and includes the likelihood with which an insect belongs to this class.
- Six insects considered
- Insect sounds in wave format
- Collected from Internet
- Online feature extraction & classification
- Direct comparison between ground truth and recognised result
- Playback of insect sounds
- Hold out accuracy on 4,500 sounds of 5 insects: 0.9056
- Holdout accuracy on 45,000 sounds of 10 insects: 0.616
- Final score (average): 0.7608
- Advanced machine learning method to optimize the parameters of insect sound classification
- Automatic insect classification by integrating contextual sensor information