Learning Domain-Robust Bioacoustic Representations for Mosquito Species Classification with Contrastive Learning and Distribution Alignment

We utilize mosquito audio datasets recorded in four different countries and multiple regions to create a multi-domain mosquito dataset comprising of eight species: An Arabiensis,Culex Pipiens, Ae Aegypti, An Funestus, An Squamosus, An coustani, Ma Uniformis, Ma Africanus, Ae Albopictus. The first seven species are from the HumBug dataset [1] recorded in Tanzania, denoted as Domain 1 (D1), comprising 37688 audio clips, totalling about 20.94 hours. The Kasetsart dataset (denoted D2) of Ae Albopictus recorded in Thailand, which is part of the HumBug project [5], contains 655 audio clips, totalling about 0.37 hours. The UFRGS dataset [10] (denoted as D3) of male and female mosquito Ae Aegypti and Ae Albopictus recorded in Brazil contains 16727 audio clips with a total duration of about 9.30 hours. The Abuzz dataset [9] (denoted as D4) of mosquito An Arabiensis, Culex Pipiens, Ae Aegypti, and Ae Albopictus recorded in the USA contains 5054 audio clips with a total duration of about 2.81 hours. The total duration of the four-domain MSC dataset used in this paper is 33.42 hours. Fig. 3 shows the distribution of wingbeat frequencies for the 8 mosquito species used in this paper, from audio clips randomly selected from the dataset. In our experiments, the duration of training, validation, and test sets is 23.46, 4.26, and 5.70 hours, respectively.

For more details, please see the paper. https://arxiv.org/pdf/2510.00346 

all_species_labels = ['an arabiensis', 'culex pipiens complex', 'ae aegypti', 'an funestus ss', 'an squamosus', 'an coustani', 'ma uniformis',  'aedes albopictus']