Unlabeled Echoes · Code for Conservation

TL;DR

Problem: Passive acoustic monitoring produces far more bat recordings than experts can label, leaving far more data unlabeled than labeled.

Idea: Simple pseudo-labeling uses unlabeled recordings effectively, while genus-aware smoothing adds biological structure to training.

Result: Strong gains across European and South African bat datasets.

Pseudo-labeling gains over the 10 percent baseline: 6.38 mixture macro-F1 points, 5.42 single-species accuracy points, and 10.98 single-species macro-F1 points.

Learning from unlabeled data

Online pseudo-labeling with interleaved labeled updates.

A supervised BAT checkpoint turns unlabeled calls into hard one-hot targets and continues training on both human labels and model-generated labels.

1Labeled calls
and human annotations

2Supervised BAT
checkpoint

3Online hard targets
for unlabeled calls

4BAT self-training on
labels + pseudo-labels

Stable self-training. The pseudo-label loss is introduced gradually, while regular human-labeled updates keep early noisy predictions from dominating training.

Datasets

Two regions, two realistic unlabeled-data settings.

Skiba simulates label scarcity through a held-out split; UFS adds a separate pool of unlabeled field recordings.

Skiba

European archive · held-out split

Source archive: >1,500 recordings, 29 species, 96 kHz
Model outputs: 18 species
Sequence split: 19,030 total: 11,186 train, 2,890 validation, 4,954 test
Supervision: 1,119 labeled training sequences (10%); 10,067 training sequences treated as unlabeled

UFS

South African archive · real unlabeled split

Source archive: >40,000 recordings, 17 species, about 15,000 without labels, 256 kHz
Model outputs: 9 bat taxa + pesticide spray
Sequence split: 15,863 labeled total: 6,736 train, 3,857 validation, 5,270 test
Supervision: 30,000 additional unlabeled field sequences

Learning across classes

Genus-aware smoothing encodes plausible confusions.

Most target weight stays on the labeled species, while a small share goes to species from the same genus.

Comparison of hard, uniform, genus-aware, and combined smoothing targets, with absolute recognition gains over hard targets.

Results

Simple pseudo-labeling is a strong baseline.

Scores are percentages. Higher is better for every metric.

Skiba dataset

10% labeled unless noted

Method	Labeled	Uses unlabeled	Mixture macro-F1	Single-species accuracy	Single-species macro-F1
Full supervision	100%	No	72.46	84.00	80.33
10% baseline	10%	No	58.74	75.06	62.47
Autoencoder	10%	Yes	44.74	68.53	53.61
BYOL	10%	Yes	52.68	69.54	57.69
Noisy Student	10%	Yes	59.71	75.22	67.49
FixMatch	10%	Yes	61.72	76.06	68.71
Pseudo-label → FixMatch	10%	Yes	63.39	76.58	69.80
FixMatch → Pseudo-label	10%	Yes	62.18	76.66	67.97
Pseudo-label	10%	Yes	65.12	80.48	73.45

UFS dataset

Independent unlabeled field pool

Method	Uses unlabeled	Mixture macro-F1	Single-species accuracy	Single-species macro-F1
Supervised baseline	No	55.25	67.55	48.65
BYOL	Yes	55.82	75.29	49.01
FixMatch	Yes	57.98	76.99	49.04
Pseudo-label	Yes	56.69	78.24	53.61

Takeaway

Unlabeled recordings can provide useful supervision without changing the model.

Online pseudo-labeling improves recognition in both datasets, and genus-aware smoothing adds a complementary biological prior.

BibTeX

@inproceedings{fundel2026unlabeled,
  title     = {Unlabeled Echoes: Pseudo-Labels and Genus-Aware
               Smoothing for Bat Call Recognition},
  author    = {Fundel, Frank and Howard, Alexandra},
  booktitle = {CV4Ecology},
  year      = {2026}
}