BioDCASE 2026 Challenge: Cross-Domain Mosquito Species Classification

The development dataset is released for the BioDCASE 2026 Cross-Domain Mosquito Species Classification task to support model development, validation, and transparent baseline reproduction. Full task information, including the challenge overview, timeline, and evaluation setting, evaluation dataset release, and submission instructions, is provided on the official task page. The fully open baseline implementation, including code and released resources, is provided through the official GitHub repository.

The released development dataset contains 271,380 audio clips in total, corresponding to 218,388.40 seconds (60.66 hours) of mosquito flight sound recordings. It covers 9 target species across 5 domains and is intended to support research on mosquito species classification under domain shift. 

The 9 target species are:
Ae. aegypti, Ae. albopictus, Cx. quinquefasciatus, An. gambiae, An. arabiensis, An. dirus, Cx. pipiens, An. minimus, and An. stephensi.

The number of clips for each species in the released development dataset is:
Ae. aegypti: 81,587
Ae. albopictus: 18,517
Cx. quinquefasciatus: 72,056
An. gambiae: 46,998
An. arabiensis: 21,117
An. dirus: 127
Cx. pipiens: 29,754
An. minimus: 550
An. stephensi: 674

The dataset spans 5 domains, with the following clip counts:
D1: 4,065
D2: 784
D3: 679
D4: 200
D5: 265,652

Each audio file follows the naming format S_<speciesID>_D_<domainID>_<clipIndex>, so both species identity and domain identity are directly accessible from the audio ID. This makes the released dataset fully transparent and easy to inspect. Participants can directly analyse species-domain distributions, reproduce the released baseline setting, or construct alternative development splits when needed.

For the released baseline, the development dataset is divided into a trainval pool of 244,163 clips and a test set of 27,217 clips. A validation set is then derived from the trainval pool by random species-stratified sampling, yielding 213,647 training clips and 30,516 validation clips. This released split is intended as a simple and reproducible reference setup for the BioDCASE 2026 Cross-Domain Mosquito Species Classification task.

The species-domain distribution is highly uneven across the development dataset. Some species-domain combinations are well represented, while others are sparse. Participants are therefore encouraged to look beyond pooled accuracy and to consider both class balance and domain balance during development.

Participants may use the species and domain information encoded in the audio IDs to construct alternative domain-aware development splits. This can help local validation better reflect the cross-domain objective of the task.

The evaluation set has now been released. It is intended only for final challenge submission and should not be used for training, validation, model selection, or hyperparameter tuning. Evaluation audio file names are randomised and do not contain species IDs, domain IDs, or seen/unseen-domain indicators. Participants should submit species predictions using the released file_id values.

The official ranking is based primarily on balanced accuracy on unseen-domain samples (BA_unseen). The domain shift gap, DSG = |BA_unseen - BA_seen|, is used as the secondary ranking metric, with smaller values preferred. BA_seen is reported for reference.

Recommended links for the page

Task page: https://biodcase.github.io/challenge2026/task5

Baseline repository: https://github.com/Yuanbo2020/CD-MSC

Official submission system: https://biodcase.github.io/challenge2026/submission

If you use the development dataset, or refer to the BioDCASE 2026 Cross-Domain Mosquito Species Classification task, please feel free to cite the following paper.

BioDCASE 2026 CD-MSC Baseline: 📄 PDF

@misc{hou2026biodcase2026challengebaseline,
      title={BioDCASE 2026 Challenge Baseline for Cross-Domain Mosquito Species Classification}, 
      author={Yuanbo Hou and Vanja Zdravkovic and Marianne Sinka and Yunpeng Li and Wenwu Wang and Mark D. Plumbley and Kathy Willis and Stephen Roberts},
      year={2026},
      eprint={2603.20118},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2603.20118}, 
}

@INPROCEEDINGS{BioL,
  author={Hou, Yuanbo and Liu, Zhaoyi and Shen, Xin and Roberts, Stephen},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Learning Domain-Robust Bioacoustic Representations for Mosquito Species Classification with Contrastive Learning and Distribution Alignment}, 
  year={2026},
  volume={},
  number={},
  pages={15207-15211},
  doi={10.1109/ICASSP55912.2026.11464393}}