Emreargin/BioDCASE2026_Bird_Counting

BioDCASE 2026 — Bird Counting (Task 6)

Development and evaluation dataset for the Bird Counting task of the BioDCASE 2026 Challenge.

📢 Evaluation set released on 1 June 2026. 10 new held-out aviaries (~380,000 audio files) are now live under eval_aviary_1/ through eval_aviary_10/. See the Evaluation set section below.

Task overview

Estimating the number of individual birds from acoustic recordings is a fundamental challenge in biodiversity monitoring. This task addresses bird abundance estimation in zoo aviaries with known ground-truth population counts.

Participants receive collections of short audio fragments (~3 seconds each) extracted from continuous passive acoustic recordings in multi-species aviaries. Each aviary contains a known number of a target bird species alongside other co-occurring species. The recordings capture birds vocalizing naturally in groups over extended periods, creating realistic acoustic complexity including overlapping vocalizations, environmental noise, and natural behavioral variation.

The task is to estimate the number of individuals of the target species in each aviary.

For full task details, timeline, evaluation criteria, and submission instructions, see:

• Task page: https://www.ml4biodiversity.org/biodcase26_birdcounts/
• Challenge page: https://biodcase.github.io/challenge2026/task6
• Baseline code: https://github.com/ml4biodiversity/biodcase-population-estimation

---

Development set

The development dataset contains 140,899 audio files across 6 aviaries recorded at European zoos using passive acoustic monitoring equipment. Recordings were made during spring and summer 2025. Each aviary was recorded continuously for 7–11 days; this dataset includes a curated subset of 2–3 representative days per aviary, selected to minimize distributional distortion of key acoustic features while keeping the dataset manageable.

Target species (development set)

Three bird species are designated as the main estimation targets for the main leaderboard:

Species	Scientific name	Dev aviaries	Population range
Greater flamingo	Phoenicopterus roseus	dev_aviary_2, dev_aviary_4, dev_aviary_5, dev_aviary_6	52–161
Red-billed quelea	Quelea quelea	dev_aviary_1, dev_aviary_3	61–153
Hadada ibis	Bostrychia hagedash	dev_aviary_2, dev_aviary_4	4–6

Each aviary also contains additional non-target bird species (2–12 species per aviary, 28 species in total across all development aviaries). The complete species inventory with population counts is provided in metadata/ground_truth.csv.

Development aviary summary

Aviary	Days	Audio files	Target species	Target population
dev_aviary_1	3	12,627	Red-billed quelea	153
dev_aviary_2	3	25,569	Greater flamingo (107), Hadada ibis (6)	113
dev_aviary_3	3	11,879	Red-billed quelea	61
dev_aviary_4	3	36,340	Greater flamingo (161), Hadada ibis (4)	165
dev_aviary_5	2	19,363	Greater flamingo	52
dev_aviary_6	3	35,121	Greater flamingo	52
Total	17	140,899

Note: dev_aviary_5 and dev_aviary_6 are two separate recording sessions from the same physical location with the same bird population, captured on different dates. They are treated as independent data points with different acoustic conditions.

---

Evaluation set

Released on 1 June 2026. The evaluation set contains ~380,000 audio files across 10 held-out aviaries recorded at European zoos using the same passive acoustic monitoring setup as the development set. Each aviary was recorded for 5–10 days; all available days are released (no per-day curation, unlike the development set).

New optional target species: Pied avocet

For BioDCASE 2026 we introduce Pied avocet (Recurvirostra avosetta) as an optional fourth target species. Pied avocets produce distinctive, well-articulated calls that contrast sharply with the synchronous flock-calling pattern of Greater flamingos — making them an interesting test case for methods that exploit individual call structure for abundance estimation.

Pied avocet predictions are not part of the main leaderboard and will not affect main rankings. They contribute to the secondary generalist leaderboard alongside non-target species predictions. Participants who only target the original three species (Greater flamingo, Red-billed quelea, Hadada ibis) are unaffected and can ignore Pied avocet entirely.

Evaluation aviary summary

Aviary	Days	Audio files	Target species
eval_aviary_1	5	21,190	Red-billed quelea
eval_aviary_2	5	15,268	Red-billed quelea
eval_aviary_3	8	50,901	Pied avocet*
eval_aviary_4	8	46,728	Greater flamingo
eval_aviary_5	10	56,092	Pied avocet*
eval_aviary_6	8	36,020	Hadada ibis
eval_aviary_7	8	46,066	Hadada ibis
eval_aviary_8	8	64,196	Greater flamingo
eval_aviary_9	7	16,658	Pied avocet*
eval_aviary_10	7	26,296	Pied avocet*
Total	74	~380,000

* Pied avocet is an additional optional target species — see the section above. Pied avocet predictions contribute only to the secondary generalist leaderboard.

Each row tells participants the target species that needs an integer count for that aviary. Each aviary also contains additional non-target species; the complete per-aviary composition and ground-truth counts for every evaluation aviary are provided in metadata/eval_ground_truth.csv, released after the challenge concluded.

Species inventory

The full per-aviary ground truth — every species and its individual count in each evaluation aviary — is available in metadata/eval_ground_truth.csv, released after the challenge concluded. During the challenge only the combined list below was published, so that the per-aviary composition (and the correspondence between some evaluation and development aviaries) stayed held-out.

A combined inventory of every species appearing somewhere in the evaluation set is also provided in metadata/eval_species_list.csv — 67 species total: 3 main target species, 1 optional target species, and 63 non-target species. No species outside this list appears anywhere in the evaluation audio, so it is the safe and tight superset to use when running pre-trained detectors with a species-filter list. The per-aviary target_species column in eval_recording_info.csv tells participants which species needs to be counted for each aviary.

Evaluation-set structure: continuation and repeated-session aviaries

The following notes document the structure of the evaluation set. Not every evaluation aviary is an independent, newly recorded environment, and this matters when interpreting per-species results.

Red-billed quelea — held-out days of development aviaries. The two quelea evaluation aviaries are not new environments. They use the remaining, previously unreleased days from the same continuous recording sessions as the corresponding development aviaries:

Eval aviary	Continues	Target population	What differs
eval_aviary_1	dev_aviary_1	Red-billed quelea (153)	Recording days only
eval_aviary_2	dev_aviary_3	Red-billed quelea (61)	Recording days only

The aviary, recording setup, and bird population are identical to the matching development aviary — only the recording days differ. For quelea, evaluation therefore measures generalization across unseen days of the same recording, not across a new environment, which is why the target counts equal the development values.

Hadada ibis — one environment recorded twice. eval_aviary_6 and eval_aviary_7 are the same physical aviary and the same bird population, captured in two separate recording sessions (different dates/times and recording setup). Both therefore have the same Hadada ibis count (4), and they should be read as two correlated recordings of a single environment rather than two independent test cases — the same situation described for dev_aviary_5 and dev_aviary_6 in the development set. This environment is new relative to the development set.

Greater flamingo and Pied avocet — new, independent environments. The two Greater flamingo aviaries (eval_aviary_4, eval_aviary_8) and the four optional Pied avocet aviaries are newly recorded environments with no development-set counterpart. The flamingo aviaries were deliberately selected to fall outside the development population range (11 and 195 individuals, versus 52–161 in development), so they test extrapolation to unseen colony sizes in addition to a new acoustic environment.

Why this matters. The meaning of "generalization" is not uniform across the target species: quelea is the same environment across unseen days, Hadada ibis is a single environment recorded twice, and Greater flamingo (and Pied avocet) require genuine cross-aviary generalization — with the flamingo aviaries additionally probing out-of-range population sizes. Aggregate scores should be interpreted with this structure in mind, and the two Hadada ibis aviaries (and the paired quelea recordings) are not statistically independent.

---

Audio format

All audio files are stereo WAV files, 16-bit PCM, sampled at 48 kHz, with a duration of approximately 3 seconds each. The files represent consecutive, non-overlapping segments extracted from continuous recordings. Format is identical between the development and evaluation sets.

Dataset structure

BioDCASE2026_Bird_Counting/
├── dev_aviary_1/                # ─── development set (with ground truth) ───
│   ├── chunk_000/
│   │   ├── rec_d1_00_00_45.750000.wav
│   │   ├── rec_d1_00_01_49.wav
│   │   └── ...
│   ├── chunk_001/
│   └── ...
├── dev_aviary_2/
│   └── ...
├── ...
├── dev_aviary_6/
│   └── ...
├── eval_aviary_1/               # ─── evaluation set (held-out) ───
│   ├── chunk_000/
│   │   └── ...
│   └── ...
├── eval_aviary_2/
│   └── ...
├── ...
├── eval_aviary_10/
│   └── ...
└── metadata/
    ├── ground_truth.csv             # development set ground truth
    ├── recording_info.csv           # development set summary
    ├── eval_ground_truth.csv        # evaluation set ground truth (per-aviary, all species)
    ├── eval_recording_info.csv      # evaluation set summary + target species per aviary
    └── eval_species_list.csv        # evaluation set combined species inventory (67 species)

Filename convention

Audio filenames follow the pattern:

rec_{day}_{HH}_{MM}_{SS}[.ffffff].wav

where {day} is a day identifier (d1, d2, ...) and {HH}_{MM}_{SS}[.ffffff] encodes the time of day. For example, rec_d1_19_05_02.500000.wav is a recording from day 1 at 19:05:02.5.

Day identifiers are anonymized — the mapping from day identifiers to calendar dates is not provided to participants. The day numbering is local to each aviary (i.e. d1 in eval_aviary_1 is unrelated to d1 in eval_aviary_2 or to any d1 in the development set).

Chunk subdirectories

Within each aviary, audio files are organized into chunk_NNN/ subdirectories for practical file management on Hugging Face (per-directory file-count limits). The chunk boundaries have no acoustic significance — they are simply a way to keep directory sizes manageable. All chunks within an aviary should be treated as a single continuous collection.

---

Metadata

metadata/ground_truth.csv (development set)

Complete species inventory for all 6 development aviaries, including both target and non-target species:

Column	Description
aviary_id	Aviary identifier (dev_aviary_1 through dev_aviary_6)
common_name	English common name of the species
scientific_name	Binomial scientific name
count	Number of individuals present in the aviary
is_target	1 if the species is evaluated for population estimation, 0 otherwise

metadata/recording_info.csv (development set)

Summary statistics per development aviary:

Column	Description
aviary_id	Aviary identifier
n_days	Number of recording days included
n_files	Total number of audio files

metadata/eval_ground_truth.csv (evaluation set)

Complete species inventory and ground-truth counts for all 10 evaluation aviaries, including both target and non-target species. Released after the challenge concluded; same format as the development-set ground_truth.csv, with an extended is_target column that distinguishes the optional target:

Column	Description
aviary_id	Aviary identifier (eval_aviary_1 through eval_aviary_10)
common_name	English common name of the species
scientific_name	Binomial scientific name
count	Number of individuals present in the aviary
is_target	0 = non-target, 1 = main-leaderboard target (Greater flamingo, Red-billed quelea, Hadada ibis), 2 = optional target (Pied avocet)

metadata/eval_recording_info.csv (evaluation set)

Per-aviary summary for the evaluation set, including the target species each aviary should be scored on:

Column	Description
aviary_id	Aviary identifier (eval_aviary_1 through eval_aviary_10)
n_days	Number of recording days included
n_files	Total number of audio files
n_chunks	Number of chunk_NNN/ subdirectories under this aviary
chunk_size	Maximum files per chunk
target_species	The target species that needs to be counted for this aviary
is_optional_target	1 if the target is the optional Pied avocet, 0 for main-leaderboard targets

metadata/eval_species_list.csv (evaluation set)

Combined inventory of every species that appears somewhere in the evaluation set, with no per-aviary breakdown:

Column	Description
common_name	English common name (sentence case)
scientific_name	Binomial scientific name
category	One of: main_target (Greater flamingo, Hadada ibis, Red-billed quelea), optional_target (Pied avocet), or non_target

---

Baseline system

A complete baseline system is available at https://github.com/ml4biodiversity/biodcase-population-estimation. It implements a two-stage pipeline:

1. Species detection — Run a bird species detector on each aviary's audio files. Two detection packages are provided:

   • pip install aria-inference (ARIA ensemble detector, recommended)
   • pip install aria-inference-birdnet (BirdNET-based detector tailored to this task)

2. Feature extraction — Extract detection-count statistics, temporal bout structure, and optionally scikit-maad acoustic indices from the detection output.

3. Population estimation — Fit species-specific regression models using leave-one-out cross-validation.

The baseline achieves a combined MAE of 11.50 (MAPE 10.6%) across all target species on the development set using ARIA detections.

Evaluation

The main leaderboard ranks systems based on population estimation accuracy for the three main target species (Greater flamingo, Red-billed quelea, Hadada ibis). The primary metric is Mean Absolute Error (MAE) computed across all required (aviary, target species) data points in the evaluation set. Secondary metrics include RMSE, R², and MAPE.

Predictions for Pied avocet (optional) and other non-target species contribute to a secondary generalist leaderboard that does not affect the main ranking.

For the full submission format and timeline, see the challenge task page.

Key challenges

• Flock-calling species: Greater flamingos vocalize synchronously in large groups, making it difficult to distinguish individual contributions from detection counts alone. Raw detection rates saturate as flock size grows.
• Sparse calibration data: With only 6 development aviaries (and 2–4 data points per target species), models must generalize from very few examples.
• Multi-species environments: Each aviary contains multiple co-occurring species with overlapping frequency ranges and calling times.
• Population range: Target populations span two orders of magnitude (4 to ~200 individuals), requiring methods that work across scales.
• Cross-aviary generalization: The evaluation set introduces new aviaries with new acoustic conditions and new non-target species mixtures. Methods that overfit to the specific acoustic context of the development set will not generalize.

---

Usage with 🤗 Datasets

from datasets import load_dataset

# Load the dataset (streams audio on demand)
ds = load_dataset("Emreargin/BioDCASE2026_Bird_Counting")

Or download directly and process locally:

# Clone with git-lfs
git lfs install
git clone https://huggingface.co/datasets/Emreargin/BioDCASE2026_Bird_Counting

# Run the baseline on the development set
cd biodcase-population-estimation
pip install aria-inference
aria-inference --input ../BioDCASE2026_Bird_Counting/dev_aviary_1/ \
               --output detections/dev_aviary_1_detections.csv
# ... repeat for dev_aviary_2 through dev_aviary_6

python feature_builder.py \
    --detections-dir detections/ \
    --audio-root ../BioDCASE2026_Bird_Counting/ \
    --output results/stage2_features.csv

python estimator.py --features results/stage2_features.csv

# Run inference on the evaluation set
for i in 1 2 3 4 5 6 7 8 9 10; do
  aria-inference --input ../BioDCASE2026_Bird_Counting/eval_aviary_${i}/ \
                 --output detections/eval_aviary_${i}_detections.csv
done

When running pre-trained detectors with a species-filter list on the evaluation set, use metadata/eval_species_list.csv as the safe superset of species that may appear. The target_species column in metadata/eval_recording_info.csv tells you which species needs a count for each aviary.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use this dataset, please cite:

@dataset{ml4biodiversity2026dataset,
  author       = {Arg{\i}n, Emre and H{\"a}rm{\"a}, Aki and Arslan-Dogan, Aysenur},
  title        = {{BioDCASE 2026 Bird Counting: Avian Population Estimation
                   from Passive Acoustic Recordings}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Emreargin/BioDCASE2026_Bird_Counting},
}

Please cite this repository if you use the official baseline implementation:

@software{ml4biodiversity2026baseline,
  author       = {Arg{\i}n, Emre and H{\"a}rm{\"a}, Aki and Arslan-Dogan, Aysenur},
  title        = {{BioDCASE 2026 Bird Counting Baseline: Avian Population Estimation
                   from Passive Acoustic Recordings}},
  year         = {2026},
  publisher    = {GitHub},
  url          = {https://github.com/ml4biodiversity/biodcase-population-estimation},
  version      = {1.0.0},
}

Contact

For questions about the dataset or the challenge task, please contact:

• Emre Argın — Maastricht University (challenge task lead)
• Aki Härmä — Maastricht University
• Aysenur Arslan-Dogan — Maastricht University (main contact person) — aysenur.arslan-dogan@maastrichtuniversity.nl

Or open a discussion on the dataset page.