Cross-Domain Mosquito Species Classification

Results

Task description

Task 5 evaluates cross-domain mosquito species classification. The evaluation set contains clips from seen and unseen recording domains. More details can be found on the task description page:

Task description page

Awards

The recognitions below are highlighted separately from the numerical rankings. The official winning submission is determined by the official evaluation ranking. The Jury Award and special recognitions are selected by the task organisers based on the submitted systems and technical reports.

Official winning submission

Oliver Gstöttenbauer, Andrii Ushakov, Artem Shestakov
Johannes Kepler University Linz, Austria

Jury award

Tianyan Deng, Wanchuan Chen, Jiahao Du, Rui Gao, Yanxiong Li
South China University of Technology, China
Domain-Balanced Representation Learning and Multi-Prototype Mahalanobis Inference for Cross-Domain Mosquito Classification
Selected for the quality of the proposed system and technical report, especially the analysis of species-domain imbalance, targeted class-conditional domain alignment, and multi-prototype inference for unseen-domain generalisation.

Best robust system

Shuanglin Li, Ruxiao Qian
XiangJiang Laboratory, China
Reliability-Guided Distillation for Cross-Domain Mosquito Species Classification
Recognised for its species-domain held-out validation protocol, reliability-guided distillation strategy, and guarded ensemble design for unseen-domain robustness and domain-gap reduction.

Special Mention for Comprehensive System Integration

Meiting Lyu, Ruicheng Liu, Xu Shen, Zongye Chen, Jiatong Wu, Ziyu Wang, Wei Liu, Gongping Huang
Wuhan University, China; Waseda University, Japan
Addressing Class Imbalance in Cross-Domain Mosquito Bioacoustic Species Classification with Balanced Multi-Task Learning
Recognised for a comprehensive engineering system combining class-balanced data preparation, domain-generalised representation learning, multi-task objectives, and confidence-gated model fusion.

Official team ranking

The official team ranking assigns one rank to each team. Each row shows the team's best valid submission only; all submitted systems are listed in the submission-level table below.

Team
rank		 Team Best submission Best system name Technical
report		 Best system
rank		 BAunseen DSG BAseen
1 Shestakov_JKU Shestakov_JKU_Task5_2 FG-harmonic agreement-gate ensemble shestakovJKUTask5 1 0.313 0.338 0.651
2 Deng_SCUT Deng_SCUT_Task5_4 Domain-Class Mahalanobis dengSCUTTask5 4 0.274 0.492 0.766
3 Eva_UL Eva_UL_Task5_2 Domain-adversarial AST evaULTask5 5 0.268 0.224 0.491
4 Zhou_UB Zhou_UB_Task5_4 Domain-bucket PANNs/AST best-overall fusion zhouUBTask5 7 0.260 0.511 0.772
5 Song_HHU Song_HHU_Task5_1 AudioTransformer_Mosquito songHHUTask5 8 0.255 0.216 0.471
6 Saha_Mila Saha_Mila_Task5_1 Combined domain-generalisation + cDANN ensemble sahaMilaTask5 10 0.242 0.399 0.642
7 Gupta_AP Gupta_AP_Task5_3 CRNN-AST fusion with fixed species-9 calibration hedge guptaAPTask5 12 0.238 0.282 0.520
8 Lyu_WHU Lyu_WHU_Task5_1 DenseNet121 with Domain-Adversarial Training and CosFace Metric Learning lyuWHUTask5 19 0.205 0.431 0.636
9 You_PKU You_PKU_Task5_4 TrainVal harmonic summary SGD margin-gated prior calibrated youPKUTask5 22 0.174 0.265 0.440
10 Baseline Baseline Baseline biodcase2026Task5Baseline 26 0.172 0.409 0.582
11 LeCallet_PolytechNantes LeCallet_PolytechNantes_Task5_1 LeCallet PolytechNantes task5 1 report lecalletPolytechNantesTask5 28 0.163 0.358 0.521
12 Li_XJLAB Li_XJLAB_Task5_4 Low-DSG guarded ensemble liXJLABTask5 29 0.157 0.379 0.535
13 Wang_HHU Wang_HHU_Task5_1 Baseline-style CNN wangHHUTask5 34 0.111 0.000 0.111

Submission-level system ranking

The submission-level system ranking lists every independently submitted system and the baseline reference. This table is sorted by BAunseen with DSG as the tie-breaker, but it does not replace the official team ranking above.

System
rank		 Team Submission System name Technical
report		 Team
rank		 Within-team
rank		 BAunseen DSG BAseen
1 Shestakov_JKU Shestakov_JKU_Task5_2 FG-harmonic agreement-gate ensemble shestakovJKUTask5 1 1 0.313 0.338 0.651
2 Shestakov_JKU Shestakov_JKU_Task5_1 Perch agreement-gate ensemble shestakovJKUTask5 1 2 0.312 0.340 0.652
3 Shestakov_JKU Shestakov_JKU_Task5_3 2-voter agreement-gate ensemble shestakovJKUTask5 1 3 0.308 0.341 0.649
4 Deng_SCUT Deng_SCUT_Task5_4 Domain-Class Mahalanobis dengSCUTTask5 2 1 0.274 0.492 0.766
5 Eva_UL Eva_UL_Task5_2 Domain-adversarial AST evaULTask5 3 1 0.268 0.224 0.491
6 Deng_SCUT Deng_SCUT_Task5_2 Single-Prototype Mahalanobis dengSCUTTask5 2 2 0.268 0.570 0.837
7 Zhou_UB Zhou_UB_Task5_4 Domain-bucket PANNs/AST best-overall fusion zhouUBTask5 4 1 0.260 0.511 0.772
8 Song_HHU Song_HHU_Task5_1 AudioTransformer_Mosquito songHHUTask5 5 1 0.255 0.216 0.471
9 Deng_SCUT Deng_SCUT_Task5_3 Mahalanobis (other seed) dengSCUTTask5 2 3 0.248 0.602 0.850
10 Saha_Mila Saha_Mila_Task5_1 Combined domain-generalisation + cDANN ensemble sahaMilaTask5 6 1 0.242 0.399 0.642
11 Deng_SCUT Deng_SCUT_Task5_1 Multi-Prototype Mahalanobis dengSCUTTask5 2 4 0.240 0.550 0.789
12 Gupta_AP Gupta_AP_Task5_3 CRNN-AST fusion with fixed species-9 calibration hedge guptaAPTask5 7 1 0.238 0.282 0.520
13 Zhou_UB Zhou_UB_Task5_3 Domain-bucket PANNs/AST best-unseen fusion zhouUBTask5 4 2 0.237 0.520 0.757
14 Eva_UL Eva_UL_Task5_1 pYIN pitch-statistics Random Forest evaULTask5 3 2 0.237 0.192 0.428
15 Gupta_AP Gupta_AP_Task5_2 CRNN shift-TTA with AST auxiliary logit fusion guptaAPTask5 7 2 0.236 0.284 0.521
16 Zhou_UB Zhou_UB_Task5_2 Sliding mean-prob domain conditional zhouUBTask5 4 3 0.228 0.380 0.607
17 Zhou_UB Zhou_UB_Task5_1 Domain-bucket ensemble zhouUBTask5 4 4 0.227 0.398 0.625
18 Gupta_AP Gupta_AP_Task5_1 CRNN with domain-adversarial training and shift TTA guptaAPTask5 7 3 0.216 0.268 0.484
19 Lyu_WHU Lyu_WHU_Task5_1 DenseNet121 with Domain-Adversarial Training and CosFace Metric Learning lyuWHUTask5 8 1 0.205 0.431 0.636
20 Gupta_AP Gupta_AP_Task5_4 Domain-Style CRNN and Mild SupCon Fusion guptaAPTask5 7 4 0.190 0.230 0.420
21 Shestakov_JKU Shestakov_JKU_Task5_4 MTRCNN contrastive shestakovJKUTask5 1 4 0.180 0.304 0.484
22 You_PKU You_PKU_Task5_4 TrainVal harmonic summary SGD margin-gated prior calibrated youPKUTask5 9 1 0.174 0.265 0.440
23 You_PKU You_PKU_Task5_3 TrainVal harmonic summary SGD prior calibrated youPKUTask5 9 2 0.174 0.289 0.464
24 You_PKU You_PKU_Task5_2 TrainVal harmonic summary SGD sqrt-prior calibrated youPKUTask5 9 3 0.174 0.297 0.472
25 You_PKU You_PKU_Task5_1 TrainVal harmonic summary SGD baseline youPKUTask5 9 4 0.174 0.304 0.478
26 Baseline Baseline Baseline biodcase2026Task5Baseline 10 1 0.172 0.409 0.582
27 Lyu_WHU Lyu_WHU_Task5_3 DenseNet121 with Domain-Adversarial Training and CosFace Metric Learning lyuWHUTask5 8 2 0.170 0.516 0.686
28 LeCallet_PolytechNantes LeCallet_PolytechNantes_Task5_1 LeCallet PolytechNantes task5 1 report lecalletPolytechNantesTask5 11 1 0.163 0.358 0.521
29 Li_XJLAB Li_XJLAB_Task5_4 Low-DSG guarded ensemble liXJLABTask5 12 1 0.157 0.379 0.535
30 Li_XJLAB Li_XJLAB_Task5_2 GRL-KD stable ensemble liXJLABTask5 12 2 0.151 0.365 0.515
31 Lyu_WHU Lyu_WHU_Task5_4 DenseNet121 with Domain-Adversarial Training and CosFace Metric Learning lyuWHUTask5 8 3 0.141 0.499 0.639
32 Li_XJLAB Li_XJLAB_Task5_3 GRL-KD guarded logit ensemble liXJLABTask5 12 3 0.135 0.370 0.504
33 Li_XJLAB Li_XJLAB_Task5_1 GRL-KD diverse guarded ensemble liXJLABTask5 12 4 0.118 0.394 0.513
34 Wang_HHU Wang_HHU_Task5_1 Baseline-style CNN wangHHUTask5 13 1 0.111 0.000 0.111
35 Lyu_WHU Lyu_WHU_Task5_2 DenseNet121 with Domain-Adversarial Training and CosFace Metric Learning lyuWHUTask5 8 4 0.067 0.582 0.649

Cross-domain performance

This view compares seen-domain and unseen-domain balanced accuracy. Points closer to the diagonal have a smaller discrepancy between seen-domain and unseen-domain performance, but the official ranking still prioritises BAunseen.

System
rank		 Submission Team Technical
report		 BAunseen BAseen DSG
1 Shestakov_JKU_Task5_2 Shestakov_JKU shestakovJKUTask5 0.313 0.651 0.338
2 Shestakov_JKU_Task5_1 Shestakov_JKU shestakovJKUTask5 0.312 0.652 0.340
3 Shestakov_JKU_Task5_3 Shestakov_JKU shestakovJKUTask5 0.308 0.649 0.341
4 Deng_SCUT_Task5_4 Deng_SCUT dengSCUTTask5 0.274 0.766 0.492
5 Eva_UL_Task5_2 Eva_UL evaULTask5 0.268 0.491 0.224
6 Deng_SCUT_Task5_2 Deng_SCUT dengSCUTTask5 0.268 0.837 0.570
7 Zhou_UB_Task5_4 Zhou_UB zhouUBTask5 0.260 0.772 0.511
8 Song_HHU_Task5_1 Song_HHU songHHUTask5 0.255 0.471 0.216
9 Deng_SCUT_Task5_3 Deng_SCUT dengSCUTTask5 0.248 0.850 0.602
10 Saha_Mila_Task5_1 Saha_Mila sahaMilaTask5 0.242 0.642 0.399
11 Deng_SCUT_Task5_1 Deng_SCUT dengSCUTTask5 0.240 0.789 0.550
12 Gupta_AP_Task5_3 Gupta_AP guptaAPTask5 0.238 0.520 0.282
13 Zhou_UB_Task5_3 Zhou_UB zhouUBTask5 0.237 0.757 0.520
14 Eva_UL_Task5_1 Eva_UL evaULTask5 0.237 0.428 0.192
15 Gupta_AP_Task5_2 Gupta_AP guptaAPTask5 0.236 0.521 0.284
16 Zhou_UB_Task5_2 Zhou_UB zhouUBTask5 0.228 0.607 0.380
17 Zhou_UB_Task5_1 Zhou_UB zhouUBTask5 0.227 0.625 0.398
18 Gupta_AP_Task5_1 Gupta_AP guptaAPTask5 0.216 0.484 0.268
19 Lyu_WHU_Task5_1 Lyu_WHU lyuWHUTask5 0.205 0.636 0.431
20 Gupta_AP_Task5_4 Gupta_AP guptaAPTask5 0.190 0.420 0.230
21 Shestakov_JKU_Task5_4 Shestakov_JKU shestakovJKUTask5 0.180 0.484 0.304
22 You_PKU_Task5_4 You_PKU youPKUTask5 0.174 0.440 0.265
23 You_PKU_Task5_3 You_PKU youPKUTask5 0.174 0.464 0.289
24 You_PKU_Task5_2 You_PKU youPKUTask5 0.174 0.472 0.297
25 You_PKU_Task5_1 You_PKU youPKUTask5 0.174 0.478 0.304
26 Baseline Baseline biodcase2026Task5Baseline 0.172 0.582 0.409
27 Lyu_WHU_Task5_3 Lyu_WHU lyuWHUTask5 0.170 0.686 0.516
28 LeCallet_PolytechNantes_Task5_1 LeCallet_PolytechNantes lecalletPolytechNantesTask5 0.163 0.521 0.358
29 Li_XJLAB_Task5_4 Li_XJLAB liXJLABTask5 0.157 0.535 0.379
30 Li_XJLAB_Task5_2 Li_XJLAB liXJLABTask5 0.151 0.515 0.365
31 Lyu_WHU_Task5_4 Lyu_WHU lyuWHUTask5 0.141 0.639 0.499
32 Li_XJLAB_Task5_3 Li_XJLAB liXJLABTask5 0.135 0.504 0.370
33 Li_XJLAB_Task5_1 Li_XJLAB liXJLABTask5 0.118 0.513 0.394
34 Wang_HHU_Task5_1 Wang_HHU wangHHUTask5 0.111 0.111 0.000
35 Lyu_WHU_Task5_2 Lyu_WHU lyuWHUTask5 0.067 0.649 0.582

Metric-specific rankings

This table makes the ranking fields explicit. BAunseen rank corresponds to the primary metric, DSG rank corresponds to the secondary metric, and BAseen rank is reported for reference.

Submission Team BAunseen
rank		 BAunseen DSG
rank		 DSG BAseen
rank		 BAseen
Shestakov_JKU_Task5_2 Shestakov_JKU 1 0.313 14 0.338 9 0.651
Shestakov_JKU_Task5_1 Shestakov_JKU 2 0.312 15 0.340 8 0.652
Shestakov_JKU_Task5_3 Shestakov_JKU 3 0.308 16 0.341 10 0.649
Deng_SCUT_Task5_4 Deng_SCUT 4 0.274 27 0.492 5 0.766
Eva_UL_Task5_2 Eva_UL 5 0.268 4 0.224 25 0.491
Deng_SCUT_Task5_2 Deng_SCUT 6 0.268 33 0.570 2 0.837
Zhou_UB_Task5_4 Zhou_UB 7 0.260 29 0.511 4 0.772
Song_HHU_Task5_1 Song_HHU 8 0.255 3 0.216 30 0.471
Deng_SCUT_Task5_3 Deng_SCUT 9 0.248 35 0.602 1 0.850
Saha_Mila_Task5_1 Saha_Mila 10 0.242 24 0.399 12 0.642
Deng_SCUT_Task5_1 Deng_SCUT 11 0.240 32 0.550 3 0.789
Gupta_AP_Task5_3 Gupta_AP 12 0.238 8 0.282 21 0.520
Zhou_UB_Task5_3 Zhou_UB 13 0.237 31 0.520 6 0.757
Eva_UL_Task5_1 Eva_UL 14 0.237 2 0.192 33 0.428
Gupta_AP_Task5_2 Gupta_AP 15 0.236 9 0.284 20 0.521
Zhou_UB_Task5_2 Zhou_UB 16 0.228 21 0.380 16 0.607
Zhou_UB_Task5_1 Zhou_UB 17 0.227 23 0.398 15 0.625
Gupta_AP_Task5_1 Gupta_AP 18 0.216 7 0.268 27 0.484
Lyu_WHU_Task5_1 Lyu_WHU 19 0.205 26 0.431 14 0.636
Gupta_AP_Task5_4 Gupta_AP 20 0.190 5 0.230 34 0.420
Shestakov_JKU_Task5_4 Shestakov_JKU 21 0.180 13 0.304 26 0.484
You_PKU_Task5_4 You_PKU 22 0.174 6 0.265 32 0.440
You_PKU_Task5_3 You_PKU 23 0.174 10 0.289 31 0.464
You_PKU_Task5_2 You_PKU 24 0.174 11 0.297 29 0.472
You_PKU_Task5_1 You_PKU 25 0.174 12 0.304 28 0.478
Baseline Baseline 26 0.172 25 0.409 17 0.582
Lyu_WHU_Task5_3 Lyu_WHU 27 0.170 30 0.516 7 0.686
LeCallet_PolytechNantes_Task5_1 LeCallet_PolytechNantes 28 0.163 17 0.358 19 0.521
Li_XJLAB_Task5_4 Li_XJLAB 29 0.157 20 0.379 18 0.535
Li_XJLAB_Task5_2 Li_XJLAB 30 0.151 18 0.365 22 0.515
Lyu_WHU_Task5_4 Lyu_WHU 31 0.141 28 0.499 13 0.639
Li_XJLAB_Task5_3 Li_XJLAB 32 0.135 19 0.370 24 0.504
Li_XJLAB_Task5_1 Li_XJLAB 33 0.118 22 0.394 23 0.513
Wang_HHU_Task5_1 Wang_HHU 34 0.111 1 0.000 35 0.111
Lyu_WHU_Task5_2 Lyu_WHU 35 0.067 34 0.582 11 0.649

Method summary

The following compact method summary is extracted from the submitted reports and metadata when available. Empty cells indicate information that was not available in a machine-readable or reliably extractable form.

System
rank		 Submission Team Technical
report		 BAunseen Features Embeddings Training or augmentation Decision rule
1 Shestakov_JKU_Task5_2 Shestakov_JKU shestakovJKUTask5 0.313 foreground-harmonic liftered-cepstrum (102-d) + background-whitened log-spectrum (257-d) Perch 2.0 (1536-d) + Bird-MAE-Large (1024-d) probe-level feature Gaussian noise + channel dropout softmax-mean seed ensemble + unanimous-3 agreement gate (foreground-harmonic voter)
2 Shestakov_JKU_Task5_1 Shestakov_JKU shestakovJKUTask5 0.312 harmonic liftered-cepstrum (102-d) + background-whitened log-spectrum (257-d) Perch 2.0 (1536-d) + Bird-MAE-Large (1024-d) probe-level feature Gaussian noise + channel dropout softmax-mean seed ensemble + unanimous-3 agreement gate
3 Shestakov_JKU_Task5_3 Shestakov_JKU shestakovJKUTask5 0.308 harmonic liftered-cepstrum (102-d) Perch 2.0 (1536-d) + Bird-MAE-Large (1024-d) probe-level feature Gaussian noise + channel dropout softmax-mean seed ensemble + 2-voter agreement gate
4 Deng_SCUT_Task5_4 Deng_SCUT dengSCUTTask5 0.274 log-mel spectrogram L2-normalized, 32-dim, single prototype, Mahalanobis distance with global covariance (shrinkage=0.1), Bayesian correction (m=10.0), domain-class balanced sampling FourierMix, MixStyle mahalanobis_global_bayesian (shrinkage=0.1, Bayesian prior=10.0)
5 Eva_UL_Task5_2 Eva_UL evaULTask5 0.268 log-mel spectrogram pre-trained Audio Spectrogram Transformer backbone SpecAugment, log-mel gain jitter, circular random time shifts DANN-AST species prediction
6 Deng_SCUT_Task5_2 Deng_SCUT dengSCUTTask5 0.268 log-mel spectrogram L2-normalized, 32-dim, single prototype, Mahalanobis distance with global covariance (shrinkage=0.1), Bayesian correction (m=10.0) FourierMix, MixStyle mahalanobis_global_bayesian (shrinkage=0.1, Bayesian prior=10.0)
7 Zhou_UB_Task5_4 Zhou_UB zhouUBTask5 0.260 8 kHz log-mel MTRCNN output, PANNs CNN14 embeddings/probabilities, AST AudioSet embeddings/probabilities PANNs CNN14 / AST AudioSet pretrained representations when used by this run; otherwise not used FreqMix and DIR-like training for MTRCNN; pretrained audio representation fusion weighted probability fusion with weights base=0.32, PANNs=0.56, AST=0.12 selected on development test balanced accuracy
8 Song_HHU_Task5_1 Song_HHU songHHUTask5 0.255 LogMelSpectrogram argmax
9 Deng_SCUT_Task5_3 Deng_SCUT dengSCUTTask5 0.248 log-mel spectrogram L2-normalized, 32-dim, single prototype, Mahalanobis distance with global covariance (shrinkage=0.1), Bayesian correction (m=10.0), alternative initialization seed=2024 FourierMix, MixStyle mahalanobis_global_bayesian (shrinkage=0.1, Bayesian prior=10.0)
10 Saha_Mila_Task5_1 Saha_Mila sahaMilaTask5 0.242 log-mel energies frequency-band-selective mixing (FBS-Mix) equal-weight softmax-probability ensemble over 53 checkpoints, argmax
11 Deng_SCUT_Task5_1 Deng_SCUT dengSCUTTask5 0.240 log-mel spectrogram L2-normalized, 32-dim, K-means sub-prototypes (K=10), soft-weighted Mahalanobis distance with global covariance (shrinkage=0.72), Bayesian correction (m=2.0), temperature 3.0 FourierMix, MixStyle mahalanobis_global_bayesian_multi_soft (shrinkage=0.72, Bayesian prior=2.0, temperature=3.0)
12 Gupta_AP_Task5_3 Gupta_AP guptaAPTask5 0.238 64-bin log-mel spectrogram for CRNN MIT/ast-finetuned-audioset-10-10-0.4593 CRNN training augmentation fixed development-selected probability fusion using 0.85 CRNN
13 Zhou_UB_Task5_3 Zhou_UB zhouUBTask5 0.237 8 kHz log-mel MTRCNN output, PANNs CNN14 embeddings/probabilities, AST AudioSet embeddings/probabilities PANNs CNN14 / AST AudioSet pretrained representations when used by this run; otherwise not used FreqMix and DIR-like training for MTRCNN; pretrained audio representation fusion weighted probability fusion with weights base=0.38, PANNs=0.62, AST=0.00 selected on development test BA_unseen
14 Eva_UL_Task5_1 Eva_UL evaULTask5 0.237 pYIN fundamental-frequency contour statistics after band-pass filtering and HPSS Random Forest species prediction from high-confidence pitch statistics
15 Gupta_AP_Task5_2 Gupta_AP guptaAPTask5 0.236 64-bin log-mel spectrogram for CRNN MIT/ast-finetuned-audioset-10-10-0.4593 CRNN training augmentation fixed development-selected probability fusion using 0.85 CRNN epoch 6 shift-TTA and 0.15 AST auxiliary classifier, followed by argmax species selection
16 Zhou_UB_Task5_2 Zhou_UB zhouUBTask5 0.228 8 kHz log-mel spectrogram, sliding-window MTRCNN probabilities, tabular log-mel domain classifier PANNs CNN14 / AST AudioSet pretrained representations when used by this run; otherwise not used FreqMix and DIR-like domain robust training augmentation multi-window mean probability aggregation followed by trained domain-conditional adjustment
17 Zhou_UB_Task5_1 Zhou_UB zhouUBTask5 0.227 8 kHz log-mel spectrogram, MTRCNN probabilities, predicted-domain conditional routing PANNs CNN14 / AST AudioSet pretrained representations when used by this run; otherwise not used FreqMix and DIR-like domain robust training augmentation per-file model inference with domain-bucket ensemble weights selected on development data
18 Gupta_AP_Task5_1 Gupta_AP guptaAPTask5 0.216 64-bin log-mel spectrogram SpecAugment softmax prediction averaging over temporal shifts [-8, 0, 8],
19 Lyu_WHU_Task5_1 Lyu_WHU lyuWHUTask5 0.205 log-mel spectrogram DenseNet121 ImageNet-pretrained weights (conv0 adapted to single-channel input)
20 Gupta_AP_Task5_4 Gupta_AP guptaAPTask5 0.190 64-bin log-mel spectrogram domain-style transfer during development training fixed 0.70 domain-style epoch-3 and 0.30 mild-SupCon epoch-3 probability fusion with shifts [-8, 0, 8]
21 Shestakov_JKU_Task5_4 Shestakov_JKU shestakovJKUTask5 0.180 log-mel spectrogram (64 mels, 0-4 kHz) random 2 s crop (training only) argmax over MTRCNN species logits (single model, last-epoch checkpoint)
22 You_PKU_Task5_4 You_PKU youPKUTask5 0.174 340-dimensional harmonic summary flight-tone features Raw argmax except low-margin clips use stronger inverse-prior
23 You_PKU_Task5_3 You_PKU youPKUTask5 0.174 340-dimensional harmonic summary flight-tone features Argmax after adding a mild inverse TrainVal class-prior bias.
24 You_PKU_Task5_2 You_PKU youPKUTask5 0.174 340-dimensional harmonic summary flight-tone features Argmax after adding a mild square-root inverse TrainVal class-prior
25 You_PKU_Task5_1 You_PKU youPKUTask5 0.174 340-dimensional harmonic summary flight-tone features Raw classifier argmax.
26 Baseline Baseline biodcase2026Task5Baseline 0.172 64-bin log-mel spectrogram random 2 s crop during training species argmax from MTRCNN classifier
27 Lyu_WHU_Task5_3 Lyu_WHU lyuWHUTask5 0.170 log-mel spectrogram DenseNet121 ImageNet-pretrained weights (conv0 adapted to single-channel input)
28 LeCallet_PolytechNantes_Task5_1 LeCallet_PolytechNantes lecalletPolytechNantesTask5 0.163
29 Li_XJLAB_Task5_4 Li_XJLAB liXJLABTask5 0.157 CQT random 2-second training crops; full-clip inference without test-time crop truncation full-clip species prediction followed by logit-level weighted ensemble
30 Li_XJLAB_Task5_2 Li_XJLAB liXJLABTask5 0.151 CQT random 2-second training crops; full-clip inference without test-time crop truncation full-clip species prediction followed by probability-level weighted ensemble
31 Lyu_WHU_Task5_4 Lyu_WHU lyuWHUTask5 0.141 log-mel spectrogram DenseNet121 ImageNet-pretrained weights (conv0 adapted to single-channel input)
32 Li_XJLAB_Task5_3 Li_XJLAB liXJLABTask5 0.135 CQT random 2-second training crops; full-clip inference without test-time crop truncation full-clip species prediction followed by logit-level weighted ensemble
33 Li_XJLAB_Task5_1 Li_XJLAB liXJLABTask5 0.118 CQT random 2-second training crops; full-clip inference without test-time crop truncation full-clip species prediction followed by logit-level weighted ensemble
34 Wang_HHU_Task5_1 Wang_HHU wangHHUTask5 0.111 128-bin log-mel spectrogram argmax species prediction
35 Lyu_WHU_Task5_2 Lyu_WHU lyuWHUTask5 0.067 log-mel spectrogram DenseNet121 ImageNet-pretrained weights (conv0 adapted to single-channel input)

Technical reports

BioDCASE 2026 Challenge Baseline for Cross-Domain Mosquito Species Classification

Yuanbo Hou, Vanja Zdravkovic, Marianne Sinka, Yunpeng Li, Wenwu Wang, Mark D. Plumbley, Kathy Willis, Stephen Roberts
BioDCASE 2026 CD-MSC organising team

Method summary
Features 64-bin log-mel spectrogram
Training or augmentation random 2 s crop during training
Decision rule species argmax from MTRCNN classifier
Code
Source code

Domain-Balanced Representation Learning and Multi-Prototype Mahalanobis Inference for Cross-Domain Mosquito Classification

Tianyan Deng, Wanchuan Chen, Jiahao Du, Rui Gao, Yanxiong Li
South China University of Technology

Method summary
Features log-mel spectrogram
Embeddings L2-normalized, 32-dim, single prototype, Mahalanobis distance with global covariance (shrinkage=0.1), Bayesian correction (m=10.0), domain-class balanced sampling
Training or augmentation FourierMix, MixStyle
Decision rule mahalanobis_global_bayesian (shrinkage=0.1, Bayesian prior=10.0)

Mosquito Species Classification via Domain-Adversarial Fine-Tuning of the Audio Spectrogram Transformer

Eva Bones, Matija Marolt
University of Ljubljana, Faculty of Computer and Information Science

Method summary
Features pYIN F0 contour statistics; log-mel spectrogram
Embeddings pre-trained Audio Spectrogram Transformer backbone for System 2
Training or augmentation SpecAugment, log-mel gain jitter, circular random time shifts for the AST system
Decision rule System 1: Random Forest pitch-statistics classifier; System 2: DANN-AST species prediction

CRNN-AST fusion with fixed species-9 calibration hedge

Anant Gupta, Gourav Gupta, Aditya Gupta
Arogya Pandit

Method summary
Features 64-bin log-mel spectrogram for CRNN
Embeddings MIT/ast-finetuned-audioset-10-10-0.4593
Training or augmentation CRNN training augmentation
Decision rule fixed development-selected probability fusion using 0.85 CRNN
Code
Source code

Cross-Domain Mosquito Species Classification with Input-Level Mixup

Ewen Le Callet
Polytech Nantes

Reliability-Guided Distillation for Cross-Domain Mosquito Species Classification

Shuanglin Li, Ruxiao Qian
Xiangjiang Laboratory

Method summary
Features CQT
Training or augmentation random 2-second training crops; full-clip inference without test-time crop truncation
Decision rule full-clip species prediction followed by logit-level weighted ensemble

Addressing Class Imbalance in Cross-Domain Mosquito Bioacoustic Species Classification with Balanced Multi-Task Learning

Meiting Lyu, Ruicheng Liu, Xu Shen, Zongye Chen, Jiatong Wu, Ziyu Wang, Wei Liu, Gongping Huang
Wuhan University; Waseda University

Method summary
Features log-mel spectrogram
Embeddings DenseNet121 ImageNet-pretrained weights (conv0 adapted to single-channel input)

Combined domain-generalisation + cDANN ensemble

Sulagna Saha, Aaron Elcheson
Mila -- Quebec AI Institute, Universite de Montreal; Mila / unaffiliated

Method summary
Features log-mel energies
Training or augmentation frequency-band-selective mixing (FBS-Mix)
Decision rule equal-weight softmax-probability ensemble over 53 checkpoints, argmax

FG-harmonic agreement-gate ensemble

Oliver Gstoettenbauer, Andrii Ushakov, Artem Shestakov
Johannes Kepler University Linz

Method summary
Features foreground-harmonic liftered-cepstrum (102-d) + background-whitened log-spectrum (257-d)
Embeddings Perch 2.0 (1536-d) + Bird-MAE-Large (1024-d)
Training or augmentation probe-level feature Gaussian noise + channel dropout
Decision rule softmax-mean seed ensemble + unanimous-3 agreement gate (foreground-harmonic voter)
Code
Source code

AudioTransformer_Mosquito

Hangyu Song
Hohai University

Method summary
Features LogMelSpectrogram
Decision rule argmax

Baseline-style CNN

Junyuan Wang
Hohai University

Method summary
Features 128-bin log-mel spectrogram
Decision rule argmax species prediction

TrainVal harmonic summary SGD margin-gated prior calibrated

Yuhuan You
Peking University

Method summary
Features 340-dimensional harmonic summary flight-tone features
Decision rule Raw argmax except low-margin clips use stronger inverse-prior

Domain-bucket PANNs/AST best-overall fusion

Haoyang Zhou
University of Bremen

Method summary
Features 8 kHz log-mel MTRCNN output, PANNs CNN14 embeddings/probabilities, AST AudioSet embeddings/probabilities
Embeddings PANNs CNN14 / AST AudioSet pretrained representations when used by this run; otherwise not used
Training or augmentation FreqMix and DIR-like training for MTRCNN; pretrained audio representation fusion
Decision rule weighted probability fusion with weights base=0.32, PANNs=0.56, AST=0.12 selected on development test balanced accuracy
Code
Source code