MosquitoNet: A Siamese CNN Mosquito Species Identification System Using Audio Signals

Mosquito species classification plays a crucial role in monitoring and controlling mosquito-borne diseases. An accurate mosquito species identification system can significantly enhance public health outcomes. In this study, we introduce MosquitoNet, a Siamese Convolutional Neural Network (Siamese CNN) that classifies mosquito species based on wingbeat sounds. Our approach employs a Siamese CNN that accepts two spectrograms as inputs. Experiment results show that the Siamese CNN model trained with a categorical focal loss function, SpecMix augmentation method, undersampling, and oversampling, and utilizing the spectrogram pair, Log-mel, and MFCC as inputs, achieved the best performance with an accuracy of 90.94% and a macro-F1 score of 75.43%. MosquitoNet outperforms single-input CNN models and the state-of-the-art model, MozzBNNv2, in terms of ROC AUC and PR AUC. Notably, it demonstrates a 4.90% improvement in ROC AUC and a 16.40% improvement in PR AUC, respectively. Essentially, when evaluated using the HumBugDB dataset, our proposed model outperforms non-Siamese models that classify mosquito species using a single spectrogram as input.