Machine learning classifier to differentiate the hissing behavior from the other behavior of Apis cerana

The honeybee is one of the social insects that always communicate with its nestmates throughout all activities. They use the dance language to communicate the distance and direction of food sources. Bees integrate both visual and acoustic communication. Honeybees produce sounds during the dance which are potentially important in the transmission of information for orientation and the dance circuit of recruited bees. Dancers produce airborne sound signals during their take-off to new nest sites. Sound has also been used in the communication among the communication of disturbance via hissing communication. The monitoring of beehive has been applied in numerous publications to gather audio data and automatically classify the potential beehive status. Here we present the technique of a classification model of hissing behavior in Asian cavity-nesting bees, Apis cerana under different circumstances. We installed the monitoring device to collect hissing signal from the defensive behavior in the Asian honeybee, Apis cerana. This paper explores popular feature extraction techniques in audio processing, including spectral transformation, Mel filter banks, and Mel spectrogram. Moreover, we experimented in adjusting both the structure and trainable parameters of the one-dimensional and two-dimensional convolutional neural networks with two competing objectives: the minimum number of trainable parameters and 95% baseline accuracy. However, the random forest model, the classical machine learning model, was also tested in this study. The results revealed that the onedimensional neural network trained with the temporal domain spectrogram that consists of 2 hidden layers, 32 nodes for each layer and a minimum of 3,737 trainable parameters can provide the best accuracy. This result implies the practical application of the use of machine learning to detect hissing behaviors