Bioacoustics and Environmental Monitoring

Twelve microphone models went into the field together. The point was to measure what the microphone alone does to a wildlife survey. The 2020 result is the whole lesson in two sentences: “Microphone signal-to-noise ratio positively affects the sound detection space areas, which increased by a factor of 1.7 for audible sound, and 10 for ultrasound, from the lowest to the highest signal-to-noise ratio microphone. Consequently, the sampled vocalisation activity increased by a factor of 1.6 for birds, and 9.7 for bats.”

Nothing in those factors is about animals. Between the lowest and the highest signal-to-noise ratio microphone the birds did not become more numerous, more talkative or more active. The recorded bird activity rose by 1.6 and the bat activity by 9.7 because the sound detection space around the sensor had grown — by 1.7 for audible sound, by 10 for ultrasound. A project that swaps its recorders between seasons and then reports a rise has bought that rise from a supplier.

The obvious safeguard would not catch it. A review that only checked detection scores across different sites and seasons would see the same rise wherever the new hardware went. The extra detections come from the equipment. The equipment travels with the survey.

Between an animal and a line in a detection log sit seven separate conditions. The microphone test changed one of them without going near a bird. Bioacoustic systems detect, classify, localize, or summarize animal and environmental sounds. Whether any particular call reaches a file depends on whether the animal is there, whether it calls, how the sound travels, the weather, whether the sensor is working, where somebody put it, and how sensitive the detector is. Only the first of those is the thing an ecologist wants to know about. Move the last one and, in the microphone study's words, “the sampled vocalisation activity increased by a factor of 1.6 for birds, and 9.7 for bats.”

That is the operational boundary, and it is worth stating flatly. Detection counts are not direct abundance estimates. A silent species may be present. One individual may call repeatedly. Recording conditions can change detectability across space and time. The defensible version of the work starts by defining the ecological estimand — the quantity in the world you are trying to estimate — then models observation bias against it, and says out loud who owns each end.

BirdNET's “a F0.5 score of 0.414 for annotated soundscapes” and a rising detection count answer two different kinds of question, and only one kind is ecological. Call detection finds occurrences of an acoustic pattern and claims nothing beyond that. An F0.5 on annotated soundscapes measures precisely that, and measures it honestly.

Occupancy inference and a soundscape index are ecological claims, and they are not equally well supported. Occupancy has an estimator and a detection probability attached to it. A soundscape index has a correlation. A 2022 meta-analysis pooled 364 effect sizes from 34 studies covering the 11 most used acoustic indices. Against diversity metrics it got r = 0.33, CI [0.23, 0.43]. The evidence an ecological claim needs is ecological. That is why no detection score, however good, can be handed over in its place.

Those ecological claims get made in four words, and each of the four has a hard definition that a detector count does not meet. Occupancy, detectability, soundscape, and acoustic index each carry their own evidence, their own unit, and their own limit on what may be concluded. Borrowing one of them to describe a detector’s output is how a count starts sounding like a finding.

Hardware and placement are not the only conditions that move. Weather and sensor uptime shift a detection rate as effectively as a new microphone does, and they shift it week to week rather than once. They belong inside the observation model as terms, not in the discussion afterwards as caveats. A caveat cannot be subtracted from a count.

Spotted owls and barred owls put a number on what that costs. A 2019 study reported: “Based on ~6-night passive-acoustic surveys, site occupancy and detection probabilities were 0.43 and 0.50, respectively, for the common but declining species (the spotted owl), and 0.09 and 0.67, respectively, for the rare but increasing competitor (the barred owl).” A detection probability of 0.50 says that roughly half the six-night surveys at occupied spotted owl sites came back with nothing in them. The two species also differ: 0.50 against 0.67. The same recording effort sees them at different rates, so a comparison between their counts is a comparison between two detection probabilities unless the model says otherwise.

The design consequence is where the numbers bite hardest. Simulations showed what it takes to detect a 2% annual decline in spotted owl occupancy within 10 years with more than 80% power: 1,000 sites surveyed three times per season, or 1,500 sites surveyed twice. That is the real price of the trend claim, and it is set before any audio is recorded. The path below keeps measurement, modeling, decision, and verification visible as separate steps rather than collapsing them into one number. That is what lets you point at the step that failed when the count and the ecology disagree.

Event precision and recall under independent sites and seasons is the right figure to ask for first. What it cannot do on its own is tell you whether the detector is keying on background noise or on the recorder itself. A detector that has learned the hardware takes that advantage with it into every held-out site. The Bird Audio Detection challenge showed the cheap countermeasure: score per site instead of pooling. Pooled, one system finished second out of thirty teams. Scored on average per-site AUC, the same 'cakir' system would have ranked tenth.

So report what the number counts and which sites and seasons it covers. Report how uncertain it is, and what the weather and the recorders were doing while it was measured. Report it alongside occupancy or activity estimates that carry detection uncertainty. “Site occupancy and detection probabilities were 0.43 and 0.50, respectively, for the common but declining species (the spotted owl)”, as the owl study put it, is what a reportable pair looks like: the ecological quantity and the observation quantity, side by side, neither standing in for the other.

A monitoring study is defensible when another team can take it apart and still not find a shortcut in the background noise or the recorders. The artifact that makes that possible is three lines long. Write down what defining the ecological estimand assumes. Write down one counterexample. Write down what modeling observation bias then requires.

For a project comparing counts across seasons, those three lines are short. The assumption is that detections are comparable from one season to the next. The counterexample is 12 microphone models tested in the field, across which sampled bird vocalisation activity varied by a factor of 1.6 and bat activity by 9.7. The requirement is that hardware sensitivity and sensor placement enter the observation model as terms, so that a change in either is corrected for rather than read as birds.

The last decision is not statistical at all. It is about who else reads the coordinates. “Poaching has been documented in species within months of their taxonomic description in journals”, Lindenmayer and Scheele wrote in Science in 2017, and “over 20 newly described reptile species have been targeted in this way, potentially leading to extinction in the wild.” Their answer is a risk-tiered rule rather than a blanket one: withhold location data outright for high-value species, buffer spatial data to broad locations at moderate risk, and publish unrestricted only where the risk of perverse outcomes is low. A detection log for a rare species is a coordinate list. Which of those three tiers it falls into is a decision the study makes on purpose or by accident.