Evaluation of Bird Detection using Time-lapse Images around a Wind Farm

Introduction Environmental concerns in developing wind farms have been highlighted by both the wind-energy community and ecological experts [1–3] as the demand for wind power energy grows rapidly around the world to meet public policies for renewable energy. One of the primary concerns is the increase in bird mortality caused by collision with blades, loss of nesting and feeding grounds, and interception on migratory routes [4–7]. Hundreds of annual bird fatalities, including those of charismatic species, have been reported at several sites [6]. To assess such risks during the establishment and operation of wind farms, investigation of bird ecology and assessment of potential risks are necessary. Conventional bird monitoring has been carried out by manual observation, which is expensive and laborious [8]. Automation in this task can lower the cost, enable long-term monitoring, and lead to higher accuracy and reproducibility. However, an automatic system is required to perform bird detection as well as classification of bird species. A few studies about automatic bird monitoring exist. Although radar-based detection has been commonplace for birds [9, 10, 11], image-based detection using cameras is also a promising approach, owing to recent dramatic advances in imaging devices and the computer vision research field. DTBird [12, 13] and APEM [8, 14] are frontier enterprises developing image-based bird detection. However, very few scientific papers discuss whole pipelines designed for bird monitoring. In addition, accuracy, precision, and recall of general bird detection algorithms remain uncertain. An exception is May et al.’s work reporting that DTBird detected 76% to 96% of total birds in an experimental setting in Smøla [13]. With state-of-the-art methods in computer vision, bird detection in a general object detection competition achieves lower scores compared with detection via persons, buses, and bikes [15, 16]. The reasons for such low scores have not yet been explored.