Researchers at Cornell Lab of Ornithology, University of Massachusetts, and University of Illinois have cracked one of bird migration's deepest mysteries: how to tell what species are flying overhead when weather radar detects only shapes in the night sky. The breakthrough arrives through an unlikely partnership between citizen scientists and artificial intelligence, and it's already reshaping how we monitor, protect, and understand the journeys of more than 150 migratory bird species across North America.
For decades, weather radar has revealed the sheer volume of birds migrating—the ghostly blips that light up screens during spring and fall—but it could never answer the most essential question: which species are moving? This gap left conservationists blind to species-specific threats and migration patterns. Now, combining more than 2 billion bird observations submitted by eBird citizen scientists with data from GPS trackers, Motus radio telemetry, and banding records, a new method called BirdFlow Migration Traffic Rate (BMTR) has made species-level tracking possible across the continent.
The innovation works by teaching AI models to recognize patterns in citizen science observations, then validating those predictions against 28 years of radar data from 152 weather surveillance stations. The accuracy was striking: the models showed strong correlations with what radars were actually detecting. Researchers also compared their predictions directly against real GPS-tracked birds, confirming that the migration routes predicted by BirdFlow matched where birds actually flew. The team produced detailed migration models for 153 bird species spanning 14 orders and 39 families—a scope that would have been unimaginable a few years ago.
"BirdFlow opens up exciting new directions for monitoring and forecasting bird migration in real time," said Adriaan Dokter, project leader for BirdCast and BirdFlow at Cornell Lab. "The new BMTR metrics allow us to estimate the most likely species responsible for the movements we detect with radar, which detects the numbers of birds migrating aloft but not which species." What makes this particularly powerful is that it works even in areas where radar coverage is patchy, filling gaps that have long frustrated conservation efforts.
The practical applications are immediate and urgent. During peak migration seasons, collisions with windows kill hundreds of millions of birds annually—but different species face different risks. By knowing which birds are migrating when, building operators and conservationists can time their interventions precisely. The method is already being deployed to track avian influenza transmission routes, helping agencies monitor disease spread through waterfowl populations. Conservation planners can now study how birds within the same species use different routes, face different threats, and adapt to shifting environmental conditions—knowledge essential for designing effective protection strategies.
The research team has expanded the BirdFlow model collection from four to sixty vetted models now available through BirdFlowR software, and they envision integrating their species-specific data into existing systems like BirdCast, which already provides migration forecasts but until now lacked the ability to identify species. Perhaps most tantalizing: the methodology could eventually scale globally, depending on citizen science data availability, offering hope for tracking bird movements worldwide.
What began as a conversation between radar technicians and bird watchers has become something far more: a new way of seeing migration itself, weaving together individual tracking data, thousands of citizen scientists, and machines learning to read the night sky.