Automatic stridulation identification of bark beetles based on MFCC and BP Network.

The acoustic signals produced by insects are species-specific, and therefore can be employed for detection and identification purposes. This experiment aims to find a new way to automatically identify insects by their acoustic characteristics. Stridulation signals of four bark beetles, Dendroctonus valens, Tomicus yunnanensis, T. brevipilosus and D. armandi, were recorded under stress condition. Then after noise reduction, the signals were cut into one-echeme segments by audio software Adobe Adition2.0. With MATLAB7.1, borders of all of the segments were detected and twelve Mel-frequency cepstral coefficients were extracted as feature parameters which were then put into Back-Propagation Network as training and test samples. The numbers of training samples were set to 20, 40, 60, 80 and 100, and the number of test samples of the four beetle species were 54, 95, 54 and 50 separately. The results indicated that the identification success increased with the rise of the number of training samples. As the number of training samples increased to 100, the identification success reached the highest level, 98.14%, averaging 93.29%. In order to verify how the number of beetle species affects to recognition rates, the four species were subjected to pairwise comparison and the results showed that the recognition rates were higher than that when four species were tested together.