Avian and wildlife population surveys can aid in informing regulatory decisions, environmental assessments, and impact analyses of offshore energy development projects. Additionally, low-flying ocular aerial surveys have historically been used to estimate waterfowl populations, but place agency personnel at risk of injury and survey results are prone to bias and misclassification. The U.S. Geological Survey (USGS), in collaboration with the Bureau of Ocean Energy Management (BOEM) and the U.S. Fish and Wildlife Service Division of Migratory Bird Management (USFWS-DMBM), is advancing the development of deep learning algorithms and tools to automate the detection, enumeration, and classification of seabirds, waterfowl, and other marine wildlife. High resolution aerial imagery collected from the Atlantic Outer Continental Shelf and the Great Lakes will provide data for algorithm development. An annotation database referencing the imagery is being developed that contains targets (i.e., birds and other marine wildlife) and other relevant attributes (e.g., species, age, sex, and activity). OpenCV’s Computer Vision Annotation Tool (CVAT) is providing the framework for an interactive GUI, allowing wildlife experts to efficiently create annotations and support annotation database development. Our research also explores the impacts of image glare and background color on the performance of object detection models.
Results/Conclusions
Using our interactive GUI, we labeled approximately 30,000 bird objects in 169 images. Of these, approximately 10,000 bird objects were annotated with attributes of species, age, sex, and activity by trained biologists. Using the 10,000 annotated bird objects, we developed an object detection model using a MaskRCNN framework. The model explicitly detects birds without any further classification. Current performance is benchmarked at a mAP (mean average precision) of 0.45 and an AR (average recall) of 0.56. Clustering images by background color also affects model performance, with images of lighter blue tones performing better compared to images of darker values. Images with an increased presence of glare performed slightly better in our dataset, as mAP and AR were 0.6 and 0.7, respectively, for images with 0 – 0.1% glare while mAP and AR were 0.66 and 0.73, respectively, for images with > 1% glare. Further work will be done to implement a multi Abstractobject detection model to detect and predict objects of certain species, age, sex, and activity while taking other co-variates like ground sampling distance into account. Lastly, we find that using deep learning algorithms to detect birds reduces the time manually annotating bird objects by 95%, rapidly accelerating annotation development.