Lab 6 · BirdNET on a Dawn Chorus

Biological motivation

Passive acoustic monitoring is now the cheapest way to survey a forest. BirdNET makes it feasible to process months of recordings in days. But raw BirdNET output is not a species list — it's a list of candidate detections that need to be calibrated against ground truth before any inference. This lab teaches you that calibration step.

Data

Provided in data/dawn_chorus/:

• recordings/ — six 4-hour WAVs from a Coast Range site, recorded 04:30–08:30 on consecutive June days (24 h total).
• truth_labels.csv — 400 hand-labeled 3-second clips drawn from the same audio, stratified across the day. Columns: file, start_s, end_s, species_or_noise.
• site_meta.txt — lat/lon/elevation/week.

Tasks

Task 1 — run BirdNET (3 pts)

Run run_birdnet.py on the six recordings, using the site's lat/lon/week, with min_conf=0.1 (intentionally low so we have plenty to calibrate). Save to birdnet_detections.csv.

Task 2 — species accumulation curve (4 pts)

For each of the 24 1-hour windows, compute the cumulative number of distinct species detected at confidence ≥ 0.5, ≥ 0.7, and ≥ 0.9. Plot all three curves on one figure (species_accumulation.png).

Task 3 — calibrate against truth (8 pts)

Join BirdNET detections to the hand-labeled clips by overlapping time intervals. For each of the 6 most-common species in your truth set:

• Sweep confidence threshold 0.1 → 0.9 in steps of 0.05.
• Compute precision and recall at each threshold (true positive = clip's truth includes the species and BirdNET predicted it).
• Plot a 2x3 panel of P-R curves (calibration.png).
• Pick the species-specific threshold that gives precision ≥ 0.9, and report the corresponding recall.

Task 4 — interpretation (5 pts)

In 250 words: which species need a high threshold to be trustworthy, and which can be trusted at lower thresholds? Are there species where no threshold gives precision ≥ 0.9? What would you do for those species in a manuscript?

Go-further

Repeat the calibration restricting to detections within 30 min of sunrise. Does BirdNET get sharper during peak vocalization, or noisier?

Rubric