If you go outside at midnight on a clear April evening and remain motionless long enough, you may be able to hear it: a faint, high-pitched chip floating down from the darkness above. Then one more. Then a dozen more, hardly audible, drifted by like distant radio static. They are thrushes, warblers, and sparrows. Invisible and leisurely, hundreds of millions of them moved northward through the night sky, following paths their ancestors had charted long before cities existed below.
We were largely unaware that this was taking place for the majority of human history.
The first significant clue was discovered in 1822 when a stork—later dubbed the pfeilstorch, or arrow stork—landed in a German village with a Central African spear embedded in its neck. The fact that birds were traveling across continents while everyone was asleep was confirmed in a direct and somewhat depressing way. By the middle of the 20th century, researchers were counting the silhouettes of passing birds while pointing telescopes at full moons. Technically, it worked. It scaled horribly as well. If the sky cooperated, you could study in one place, one night.
Radar was what altered everything, first gradually and then abruptly.
Birds frequently appear in the continental radar network that the National Weather Service has been operating for decades. The returns on those radar screens fill with something other than rain each spring and fall migration. Biologists were aware of this. The volume of data was the issue. More than 200 million photos and hundreds of terabytes of data had been scanned over the course of more than 20 years. It would take more ornithologist-hours than anyone had to manually analyze even a small portion of it. No one could fully read the data, so it just sat there like a library.
Dan Sheldon, an AI researcher at the University of Massachusetts Amherst, set out to find a solution to that issue. Together with colleagues from the Cornell Lab of Ornithology, he developed a tool called MistNet, which is elegantly named after the almost invisible mist nets used by ornithologists to capture songbirds in the field. MistNet is a convolutional neural network that has been trained to distinguish between precipitation and birds in a radar image. Keep the birds and take away the rain. It sounds incredibly easy. Years passed before it was perfected.
Instead of being a research tool, what MistNet unlocked was more akin to a migration atlas. All of a sudden, researchers were able to access 24 years’ worth of stored radar data and begin asking previously unattainable questions. Where is the highest level of migration?
It turns out that the solution is a corridor that roughly runs along and just west of the Mississippi River, something that ornithologists had suspected but were never able to verify at this scale. When is migration concentrated? Surprisingly, only 10% of nights in the year see more than half of all nocturnal bird migration throughout the continental United States. There are significant ramifications to that type of statistical compression. You can begin taking action on the nights that are most important to you.
The Cornell Lab’s BirdCast platform has turned that insight into a useful tool. BirdCast, which was first introduced in 1999 but was rebuilt and significantly automated beginning in 2018, now forecasts bird migration intensity at the county level up to three days in advance using weather radar data, historical migration records, and machine learning. It monitors traffic rates, altitude, speed, and flight direction. When a significant migration wave is about to occur, it notifies conservation organizations, building managers, and city officials.
Light pollution has been the most obvious application. Artificial light is a major contributing factor to the over a billion bird deaths in the United States each year from building collisions; it disorient migrants, draws them toward cities, and keeps them circling around illuminated structures until they crash. In dozens of cities, BirdCast’s forecasts are now directly incorporated into Lights Out campaigns, providing organizers with the precise nights when turning off unnecessary lighting would have the biggest impact.
During migration season, the New York City Audubon uses BirdCast every night to locate its volunteer teams. These teams walk the city before dawn in search of injured birds near glass-covered buildings. They bring more supplies and spend more time outside during busy migration nights.
More recently, BirdCast data was directly integrated into a street lighting management platform by a company called Photometrics AI. Using real-time forecast data instead of human judgment, the system allows cities to automatically dim residential streetlights during peak migration events. Perhaps it’s insignificant compared to a billion bird deaths. However, little things are beginning to feel more urgent because the night sky is reportedly getting brighter by almost ten percent annually.
The picture of engineers and ornithologists going through decades’ worth of radar images and gradually teaching a machine to differentiate between a flock of warblers and a band of summer rain is worth pondering. Science doesn’t always make its presence known.
Sometimes it builds up silently in university offices and server rooms until, one season, the migration map appears out of nowhere. It is colorful, animated, updated almost instantly, and depicts the invisible moving across a sleeping continent. The birds never stopped flying. All we needed were the proper instruments to look up at last.
