Acoustic mosquito species classification works well in the laboratory where it was trained. Deploy the same model to a new recording site with different microphones, background noise, and acoustic conditions, and performance collapses. The baseline system in BioDCASE performs strongly on seen domains but degrades markedly on unseen ones.
The diagnosis is familiar from other machine learning problems but takes a specific form here. The classifier learns domain-specific recording artifacts rather than species-relevant acoustic cues. The frequency content of a mosquito's wingbeat is narrowband and relatively faint. The frequency content of the recording environment — microphone resonance, background insects, wind noise — is broadband and loud. The model latches onto the signal that discriminates most reliably within its training distribution, which is the recording setup rather than the mosquito.
This is a general failure mode disguised as a specific one. The paper frames cross-domain generalization as the central challenge, not within-domain recognition. And that framing is the insight. A classifier that achieves 95% accuracy on data from the same recording setup tells you about the recording setup's statistical signature. A classifier that achieves 70% accuracy across five different recording setups tells you about mosquitoes.
The practical consequence for disease surveillance is severe. Malaria, dengue, and Zika vectors need to be identified in exactly the environments where surveillance data doesn't yet exist. The model must generalize before the data arrives.
What counts as signal depends on what you're measuring — and what you're measuring depends on whether the domain changes.