Using Web images to train a deep neural network to detect sparsely distributed wildlife in large volumes of remotely sensed imagery: A case study of polar bears on sea ice

Abstract

Remote sensing can be a valuable alternative or complement to traditional techniques for monitoring wildlife populations, but often entails operational bottlenecks at the image analysis stage. For example, photographic aerial surveys have several advantages over surveys employing airborne observers or other more intrusive monitoring techniques, but produce onerous amounts of imagery for manual analysis when conducted across vast areas, such as the Arctic. Deep learning algorithms, chiefly convolutional neural networks (CNNs), have shown promise for automatically detecting wildlife in large and/or complex image sets. But for sparsely distributed species, such as polar bears (Ursus maritimus), there may not be sufficient known instances of the animals in an image set to train a CNN. We investigated the feasibility of instead providing ‘synthesized’ training data to a CNN to detect polar bears throughout large volumes of aerial imagery from a survey of the Baffin Bay subpopulation. We harvested 534 miscellaneous images of polar bears from the Web that we edited to more closely resemble 21 known images of bears from the aerial survey that were solely used for validation. We combined the Web images of polar bears with 6292 random background images from the aerial survey to train a CNN (ResNet-50), which subsequently correctly classified 20/21 (95%) bear images from the survey and 1172/1179 (99.4%) random background validation images. Given that even a small background misclassification rate could produce multitudinous false positives over many thousands of photos, we describe a potential workflow to efficiently screen out erroneous detections. We also discuss potential avenues to improve CNN accuracy, and the broader applicability of our approach to other image-based wildlife monitoring scenarios. Our results demonstrate the feasibility of using miscellaneously sourced images of animals to train deep neural networks for specific wildlife detection tasks.