A continental system for forecasting bird migration

Abstract

Billions of animals cross the globe each year during seasonal migrations, but efforts to monitor them are hampered by the irregularity and relative unpredictability of their movements. We developed a bird migration forecast system with continental scope by leveraging 23 years of spring observations to learn associations between atmospheric conditions and bird migration intensity. Our models explained up to 81% of variation in migration intensity across the United States at altitudes of 0-3000 m, and performance remained high when forecasting events 24-72 h into the future (68-72% variation explained). We infer that avian migratory movements across the United States frequently exceed 200 million individuals per night and exceed 500 million individuals per night during peak passage. Accurately forecasting bird migration will allow stakeholders to reduce collisions with illuminated buildings, airplanes, and wind turbines, predict movements under climate change scenarios, and engage the public.

MistNet: Measuring historical bird migration in the US using archived weather radar data and convolutional neural networks

“…The historical set includes data from stations KMOB (Mobile, AL) and KBGM (Binghamton, NY) from 1995 to 2017. These scans are drawn from an existing dataset of manually screened scans (Van Doren & Horton, ). Scans were selected from a 2.5‐hr period centred on three hours after local sunset on March 15th, April 15th, May 15th, September 1st, October 1st and November 1st for each station, resulting in 4,891 scans (spring, 2,549; fall, 2,342), and then manually classified as either ‘clear’ or ‘weather’ based on the presence or absence of precipitation within 37 .…”

“…A scan is accepted and used in the analysis if it is free from precipitation and clutter, otherwise it is rejected. The screening step is conducted either by a human (Buler & Dawson, ; Buler & Diehl, ; Buler et al, ; Farnsworth et al, ; Horton et al, ; McLaren et al, ; Van Doren et al, ) or using a machine‐learning classifier (Horton et al, ; RoyChowdhury et al, ; Van Doren & Horton, ). However, even a perfect whole‐scan classifier will miss biology that co‐occurs with precipitation.…”

“…Weather radar data can answer a wide range of important migration ecology questions. Previous studies have used weather radar data to understand patterns and determinants of nocturnal migration (Farnsworth et al, ; Gauthreaux et al, ; Kemp, Shamoun‐Baranes, Dokter, Loon, & Bouten, ; La Sorte, Hochachka, Farnsworth, Sheldon, Fink, et al, ), identify critical stopover habitat (Buler & Dawson, ; Buler & Diehl, ), locate on‐the‐ground roosting sites of birds (Bridge et al, ; Buler et al, ; Laughlin, Sheldon, Winkler, & Taylor, ; Laughlin et al, ; Winkler, ), understand flyways (Horton et al, ; Nilsson, Dokter, Verlinden, et al, ) and flight behaviour (Dokter, Shamoun‐Baranes, Kemp, Tijm, & Holleman, ; Horton et al, ; La Sorte, Hochachka, Farnsworth, Sheldon, Van Doren, et al, ), quantify demography (Dokter, Farnsworth, et al, ), document the effects of artificial light (McLaren et al, ; Van Doren et al, ) and disturbance (Shamoun‐Baranes et al, ) on migration, explore the projected implications of climate change (La Sorte, Horton, Nilsson, & Dokter, ), and forecast migration at continent scales (Van Doren & Horton, ). Researchers worldwide recognize the potential of radar data to provide new and urgently needed information about the migration ecology of birds, bats and insects, including detailed information about: routes, phenology, and mechanisms of migration; ecosystem services; the impacts of human activities and climate change on migration systems; conservation prioritization; aviation safety; and agricultural pests (Bauer et al, , ; Kelly & Horton, ).…”

“…Dokter et al () developed an algorithm to separate precipitation from biology in European C‐band radars; this was later extended to US dual polarization and S‐band data (Dokter, Desmet, et al, ; Dokter, Farnsworth, et al, ), but currently cannot fully separate precipitation from biology in historical US data. RoyChowdhury, Sheldon, Maji, and Learned‐Miller (), Van Doren and Horton () and Horton et al () trained machine learning models to automatically identify radar scans that are contaminated with precipitation. These methods are useful in that they allow automated analysis prior to the dual polarization upgrade in 2012–2013, but discard all biology in scans where precipitation occurs; our results will show this to be about 19% of the total biomass.…”

Radar aeroecology – a missing piece of the puzzle for studying the migration ecology of animals

“…Second, studying birth and death rates is difficult because conditions experienced during one annual stage may have delayed fitness consequences that appear in a subsequent stage (Senner et al 2015). This limitation is overcome by large-scale and long-term weather surveillance radar networks (Kelly and Horton 2016, Van Doren and Horton 2018, Nilsson et al 2019 (Fig. the monarch butterfly Danaus plexippus (Inamine et al 2016)).…”