Dual-branch time–frequency network with channel masking for automated insect sound monitoring toward field environments

Insect acoustic signals convey vital ecological information, supporting behaviors such as mating and territory defense. Monitoring these sounds toward field environments offers a non-invasive way to assess biodiversity and ecosystem health, yet automatic classification is hindered by short signal durations, spectral variability, and noise. We propose DB-CAFNet, a lightweight time–frequency model combining a 1D CNN and spectrogram encoder (WaveSpecNet) with a novel ChannelMask classifier that introduces channel-wise perturbations to enhance robustness. On InsectSet32, our method achieves a best accuracy of 89.47%, surpassing the previously reported model (80.21%) (He et al., 2024). Consistently strong results across five additional public datasets further demonstrate its generalization ability, with only 4.09M parameters and 2.55 GFLOPs under simulated field conditions.Its compact design facilitates deployment on resource-limited hardware, making the model a promising candidate for future edge-based monitoring workflows in field-oriented settings.Future work will incorporate real outdoor recordings to complete ecological validation in field environments. • A dual-branch network captures temporal and spectral features of insect sounds. • A channel mask classifier improves robustness via training-time channel dropout. • Three lightweight augmentations are introduced for better acoustic generalization. • Achieves 89.47% accuracy with 4.09M parameters on InsectSet32. • The lightweight model is a promising candidate for future edge-based monitoring.