The monitoring of migratory geese at known stopover sites is crucial to conserving their habitat, but counting large flocks of waterfowl usually requires skilled manpower. The use of observations from unmanned aerial vehicles (UAVs, otherwise known as drones) is a potential alternative to traditional bird counting methods. We used a multicopter-type UAV with a well-stabilized camera to count greater white-fronted geese (Anser albifrons) that seasonally roost at Lake Miyajima-numa, Hokkaido, Japan. Because the geese roost after sundown, it was necessary to determine camera settings that would enable the detection of geese on the lake in images taken under dim light conditions. The key camera setting was a very long exposure time of up to half a second, which allowed us to detect and count geese on images taken up to approximately 30 minutes after sunset. A single UAV flight could observe the entire lake from an altitude of 100 m above the water surface with little disturbance to the roosting geese.

We used a cascade classifier, a machine leaning technique, to automatically count the geese in the images. The counting accuracy ranged from −4.1 % to ＋6.1 % in four validation cases compared with manual counts on the UAV images. We conclude that the combination of UAVs and machine leaning methods can yield goose counts with an accuracy of ±15 %. The results suggest that this approach will be useful for monitoring geese or other waterfowl.