mosquito-mtrcnn-dg
A compact MTRCNN classifier that names the mosquito species in a flight-tone recording, trained for cross-domain generalization on the BioDCASE 2026 Task 5 development data. It knows 9 species.
[!WARNING] Research aid only. Not for public-health or vector-control decisions. Accuracy on recordings from unseen setups is low; see Results.
🚀 Usage
hf download jgalego/mosquito-mtrcnn-dg mtrcnn.py --local-dir .
uv run mtrcnn.py predict clip.wav
predict downloads model.pt, resamples each WAV file to 8 kHz, splits it into 2-second windows, averages their embeddings, and prints the species whose class mean is nearest in Mahalanobis distance. It takes any number of files. mtrcnn.py also holds the training and evaluation code.
Species: Aedes aegypti, Aedes albopictus, Culex quinquefasciatus, Anopheles gambiae, Anopheles arabiensis, Anopheles dirus, Culex pipiens, Anopheles minimus, Anopheles stephensi.
🏋️ Training
| Data | All 213,647 clips of the official training split of BioDCASE 2026 Challenge: Cross-Domain Mosquito Species Classification (Hou et al., Zenodo, 10.5281/zenodo.20478577, CC BY 4.0) |
| Model | MTRCNN with three temporal-resolution branches and a 32-dimensional embedding |
| Domain generalization | FourierMix and MixStyle across recording domains, a gradient-reversal domain classifier, class-conditional domain losses, supervised contrastive loss |
| Input | 2-second crops, 64-bin log-mel up to 4 kHz |
| Sampling | Balanced over species-domain cells, logit adjustment with the sampler's species priors |
| Schedule | 12 epochs, AdamW, learning rate 0.001, batch size 128, cosine decay; the last epoch is kept |
| Inference | Mahalanobis distance to class means of 20,000 training clips with one shared covariance matrix |
| Seed | 42 |
📊 Results
Balanced accuracy (%) on all 27,217 clips of the official development test split. Unseen clips (4,202) come from species-domain pairs that do not occur in training.
| This model | Seen | Unseen |
|---|---|---|
| Mahalanobis | 74.90 | 28.25 |
| Softmax | 69.32 | 25.47 |
Mean ± standard deviation over training seeds:
| Configuration | Inference | Seeds | Seen | Unseen |
|---|---|---|---|---|
| 2-second crops | Softmax | 3 | 67.04 ± 2.40 | 25.15 ± 0.53 |
| 2-second crops | Mahalanobis | 3 | 73.76 ± 1.14 | 27.74 ± 0.45 |
| 0.625-second crops, RMS-normalized, mel from 200 Hz | Softmax | 3 | 76.77 ± 0.23 | 20.57 ± 0.50 |
| 0.625-second crops, RMS-normalized, mel from 200 Hz | Mahalanobis | 3 | 78.09 ± 0.33 | 20.72 ± 0.69 |
| 2-second crops, liftered log-mel | Softmax | 3 | 74.60 ± 0.45 | 21.70 ± 0.49 |
| 2-second crops, liftered log-mel | Mahalanobis | 3 | 75.78 ± 0.96 | 21.76 ± 1.03 |
For reference, on the same split:
| System | Seen | Unseen |
|---|---|---|
| Official MTRCNN baseline, 10 seeds (task page) | 88.06 ± 1.08 | 17.51 ± 1.97 |
| Fine-tuned WavLM-Large | 23.74 | 17.25 |
| Probes on frozen Perch v2 embeddings, per seed † | 31-35 | |
| Our probes on Perch v2 embeddings, 8 seeds | 82.33 | 28.78 |
| Our probes on harmonic features, 8 seeds | 72.49 | 19.08 |
| Our probes on background-whitened features, 8 seeds | 74.21 | 18.34 |
| Equal-weight ensemble of the three probes and this model | 81.17 | 24.41 |
| Agreement-gated Perch, harmonic, BirdMAE and background-whitened ensemble † | 36.16 |
† Reported in aptemvs/mosquitoes-biodcase2026-task5 (legacy/final/) and not reproduced here. The checkpoints only accept precomputed Perch v2 and BirdMAE embeddings, the inference script depends on code that is not in that repository, and the ensemble's design was compared against this same test split, so these figures are likely optimistic.
Our probes use the precomputed features from the same dataset, with domain-balanced sampling and domain alignment losses. They did not reproduce the † figures, and averaging them with this model did not beat it, so we kept the model as is. Aedes aegypti stays at 0% for every system.
This model generalizes better than the MTRCNN baseline but gives up about 14 points of seen-domain accuracy.
Unseen recall per species, Mahalanobis:
| Species | Recall |
|---|---|
| Aedes aegypti | 0% |
| Aedes albopictus | 59% |
| Culex quinquefasciatus | 5% |
| Anopheles gambiae | 0% |
| Anopheles arabiensis | 0% |
| Anopheles dirus | 0% |
| Culex pipiens | 94% |
| Anopheles minimus | 74% |
| Anopheles stephensi | 22% |
⚠️ Limitations
- 99.4% of the training audio comes from one recording domain. Anopheles gambiae and Anopheles arabiensis have no training clips from any other domain.
- On unseen domains the model tends to predict a species that the training data has in the same domain, so several species are almost never recognized there.
- The configurations above were compared on this development test split, so the numbers are slightly optimistic for new recordings.