How Citizen Science and AI Built a Living Bird Forecast

A smartphone held up to the morning chorus is not, by itself, a biodiversity observatory. It may record a warbler. It may also record traffic, wind, a distant lawn mower or the birds that happen to interest the person holding the phone. Turning millions of such moments into a reliable picture of where species are—and where they may appear tomorrow—requires more than an identification app.

A Finnish research team has now tested a more ambitious architecture. Its system combines raw audio submitted by the public, a machine-learning bird classifier, long-term ecological records and a statistical model that updates every night. The result is what the researchers call a biodiversity “digital twin”: a living model of bird occurrence and vocal activity across Finland, refreshed as new observations arrive.

The important result is not the futuristic label. It is the evaluation. Across 89 bird species with enough app detections, adding the fresh citizen-science stream raised mean next-day discrimination from an area under the receiver operating characteristic curve, or AUC, of 0.71 to 0.77. In a separate field test using 1,185 five-minute counts performed by expert birders who did not know the model probabilities used to choose locations, mean AUC across 73 species rose from 0.62 to 0.67. Those are moderate improvements, not a solved biodiversity crisis. But they show that a continuously updated model can extract useful signal from a very large, imperfect public data stream.

Research references: [1]

From a recording app to an observing system

The project began with Muuttolintujen Kevät—Finnish for “the spring of migratory birds”—an app launched in March 2023 through a campaign with Finland’s public broadcaster Yle. Participants do not have to identify the birds they hear. They submit raw audio, and a classifier based on BirdNET, fine-tuned with expert annotations for 263 Finnish bird species, proposes detections. The analysis retained detections at a species-specific confidence threshold of at least 0.90.

By September 29, 2025, 315,609 people—about 5 percent of Finland’s population, according to the paper—had submitted at least one recording. The system had accumulated 16.3 million recordings containing 15.0 million bird detections above the study threshold. That scale matters, but the design choices matter more.

Citizen-science records often cluster where people live, when they have time and around the species they want to find. The Finnish app therefore supports three recording modes. A person can record directly; schedule one minute of audio every ten minutes for as long as twelve hours; or make a standardized five-minute recording at one of 580 permanent point-count locations established with municipalities and national parks. Each observation carries its location, time, recording type and duration. The raw audio is retained so detections can be checked or reclassified as the classifier improves.

This does not make the data unbiased. Direct recordings are still triggered by sounds that catch a participant’s attention. Some users hunt for species they have not yet logged, while others sample more systematically. The app also hears only vocal birds; it misses visual observations. What the design does is measure more of the observation process and create structured sampling alongside opportunistic submissions. That gives the forecasting model a better chance to distinguish a change in birds from a change in human behavior.

Research references: [1]

What the digital twin actually predicts

Every night, the system updates three probabilistic components. A migration model estimates whether a species should be present at a given latitude and time of year. A spatial model estimates whether a one-hectare location falls within the species’ distribution, using long-term surveys plus habitat, forest and climate variables. A detection model estimates whether a present bird would vocalize in a way that the phone system recognizes, given the season, time of day, recording mode and duration.

Multiplying those components yields the probability that a species will be detected in a particular recording. The “twin” is not a literal copy of an ecosystem. It is a continuously updated statistical representation of bird distribution, migration and vocal activity, together with the human sampling system that observes them. Its spatial grid is one hectare and its update cadence is daily.

That distinction is useful. Many AI-for-nature systems stop at classification: Is this sound a common gull or a sedge warbler? Classification is only the first link in a chain. Conservation decisions need estimates that account for where, when and how observations were gathered. The Finnish system joins recognition, ecological priors, sampling metadata and repeated updating in one pipeline.

Research references: [1]

Two tests, and what they do—and do not—show

The first evaluation was prospective within the app. For each day in 2024, the researchers updated the model with app data through the previous day and asked it to predict detections in the next day’s recordings. Across 89 species with at least 5,000 detections during 2024, mean AUC improved from 0.71 under the prior model to 0.77 after the daily update. Gains were largest for migratory species and for species the prior model predicted poorly.

That test is valuable because it respects time: tomorrow’s records were not used to make today’s prediction. But the test data came from the same app ecosystem, so shared behaviors or technical characteristics could make success look more transferable than it is.

The second evaluation was stronger on independence. During spring 2025, experienced volunteer birders conducted 1,185 five-minute point counts at preselected sites. The sites emphasized places where the prior and updated predictions disagreed, but the observers were blind to those probabilities. Across the 73 species available for comparison, the updated model’s mean AUC was 0.67, versus 0.62 for the prior model. For 53 species that could also be compared with eBird’s 2025 weekly abundance maps, the digital twin’s mean AUC was 0.67 and eBird’s was 0.62.

AUC measures how well a model ranks presences above absences. It does not prove that its probabilities are perfectly calibrated, that it predicts population abundance, or that anyone made a better conservation decision because of it. The expert test also covered one country, one taxonomic group and one month of spring. The paper’s own authors call the gain from 0.62 to 0.67 moderate and say the system may not transfer cleanly to citizen-science platforms that were not designed for real-time integration.

Research references: [1]

Why this matters for resilience

Climate change can shift seasonal timing and species ranges. Monitoring systems built around annual reports can miss fast biological changes or recognize them only after the moment for an effective response has passed. A daily updating model could, in principle, reveal a migration front arriving earlier than expected, a species appearing outside its usual range or a mismatch between birds and the habitats they depend on.

The Finnish work provides evidence for the first, narrower step: continuous public observations can improve short-horizon predictions of where audible bird species will be detected. The study does not show that the system has already prevented biodiversity loss, changed land management or improved climate adaptation outcomes. Those remain design targets—and they should be evaluated as rigorously as the forecast itself.

The architecture nevertheless offers a practical lesson for environmental AI. The highest-value system may not be the classifier with the best benchmark score. It may be the one built around an auditable observation process: retaining source material, recording sampling effort, updating a domain model, validating against independent experts and publishing data and code where privacy allows.

Research references: [1]

The governance questions travel with the technology

Scaling beyond Finland would not be a matter of copying an app. Bird communities, languages, smartphone access, connectivity, privacy expectations and the availability of expert reference data vary sharply. Areas with the largest biodiversity-monitoring gaps may also have the weakest digital infrastructure. Raw audio can capture human voices and precise locations can expose people or sensitive species. The Finnish pipeline generalizes public coordinates to a one-kilometre grid, applies coarser 1–100-kilometre obfuscation for sensitive taxa, limits exact-location access, and keeps associated audio nonpublic for privacy reasons; flagged records can be reviewed by assigned experts.

Those safeguards are part of the scientific system, not administrative details. A global version would need local consent, transparent rules for data use, ways for communities to challenge classifications and durable funding for expert review. It would also need evaluations for rare species, quiet species and regions where training data are thin. Otherwise, a tool meant to fill knowledge gaps could simply map the places and organisms that digital systems already hear best.

Research references: [1] [3]

A credible next step

The research team has archived the digital-twin and prior-model source code through Zenodo, made app-derived records available through the Finnish Biodiversity Information Facility, and provided source data with the paper. That makes replication and criticism more feasible. The project reports funding from Finnish, European, foundation and U.S. public research sources, and the authors declared no competing interests.

The next test should move from prediction to decisions. Can an updated forecast help a protected-area manager time a survey, detect an anomalous migration, or target habitat restoration earlier? Does it improve outcomes relative to existing monitoring, at an acceptable cost and without shifting control away from local communities? The field does not yet have those answers.

What it now has is a working demonstration that people without specialist identification skills can contribute to a serious ecological forecast—and that AI can be most useful when it strengthens, rather than replaces, the chain linking public observation, expert knowledge and accountable environmental action.

Research references: [1] [2] [3]