A crowdsourced model for automating airborne waterfowl surveys

Keywords: machine learning, UAS, waterfowl, Bosque del Apache, New Mexico
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 02:00 PM (Eastern Time (US & Canada)) - 2/26/2022 03:20 PM (Eastern Time (US & Canada))
Room: Virtual 5

Abstract

Waterfowl are important ecological, economic, cultural, and aesthetic resources worldwide, though under threat from a variety of anthropogenic sources. Accurate population counts are critical to adaptive wildlife management efforts. Imagery interpretation is more precise than traditional in-time counts performed on the ground and occupied aircraft; however, data volumes resulting from airborne imagery collection require machine interpretation to scale to meet wildlife managers’ needs. The data volumes also require image annotations (labels) in greater quantities than expert observers can typically produce in a cost-efficient manner. Thus, we compare the accuracy of a convolutional neural network trained by labels produced by biologists (Expert) and volunteers (Crowdsourced) in both single-class (detection only) and three-class (duck/goose/crane) identifications of waterfowl from drone images collected at Bosque del Apache National Wildlife Refuge in New Mexico. The Crowdsourced classifications had higher detection probability than the Expert classifications, labeling 88% - 99.7% of waterfowl compared to 80% in the Expert set, but lower overall classification accuracy (71-82% for Crowdsourced vs 98% for Expert) due overdetection of non-waterfowl objects. With a broader pool of volunteer labels, we believe that uncertainty could be reduced and higher classification accuracy achieved compared to expert-derived labels, along with ancillary benefits such as opportunities for public outreach and engagement in science.