Honey Bee Queen Presence Detection from Audio Field Recordings using Summarized Spectrogram and Convolutional Neural Networks

Abstract : The present work proposes a simple supervised method based on a downsampled time-frequency representation of the input audio signal for detecting the presence of the queen in a beehive from noisy field recordings. Our proposed technique computes a "summarized-spectrogram" of the signal that is used as the input of a deep convolutional neural network. This approach has the advantage of reducing the dimension of the input layer and the computational cost while obtaining better classification results with the same deep neural architecture. Our comparative evaluation based on a cross-validation beehive-independent methodology shows a maximal accuracy of 96% using the proposed approach applied on the evaluation dataset. This corresponds to a significant improvement of the prediction accuracy in comparison to several state-of-the-art approaches reported by the literature. Baseline methods such as MFCC, constant-Q transform and classical STFT combined with a CNN fail to generalize the prediction of the queen presence in an unknown beehive and obtain a maximal accuracy of 55% in our experiments.