Detecting polar bears on land using satellite imagery and deep learning

Abstract

Summer land use is increasing for polar bears in the Beaufort Sea, which may lead to increases in human-bear interactions. Remote methods may enable noninvasive approaches to monitor polar bear abundance and distribution on land. We examined the potential of using very high-resolution optical satellite imagery to detect polar bears on land from images collected at three sites on the Beaufort Sea coast of Alaska. Additionally, we evaluated the use of deep learning models to automate detection of polar bears in these images and the degree to which detection of polar bears differed among three very high-resolution optical satellite missions. Pléiades Neo images exhibited 90% average precision in validation analyses using a faster region-based convolutional neural network model, which was 55% and 11% greater than the average precision using WorldView-2 or -3 images, respectively. Applying the Pléiades Neo training set to new imagery exhibited lower precision (35%), but high recall (100%), which indicates that deep learning could be used to expedite the visual detection of polar bears in imagery. Very high-resolution optical satellite imagery provides a remote method for monitoring polar bears on land where they occur in high densities, but satellite mission, off-nadir angle, pansharpening methods, and image resolution are important considerations.