Instrument-based aerial surveys have proved to be an efficient approach to estimating broadly distributed populations of ice-associated seals in the Arctic by increasing survey range, decreasing disturbance of animals, and reducing error relative to surveys with human observers. We have successfully used long wavelength infra-red (IR) images to detect warm-bodied animals against the cold background of sea ice, and paired high-resolution color imagery to identify the species of detected animals. This method produces millions of images and requires automation to reduce the delay of abundance and distribution analyses. Automating data collection and image processing increases efficiency of the survey and post processing of collected imagery. We developed several deep learning models using an annotated training set of images and evaluated the models’ performance using a separate test set of images containing animals identified either by on board observers or through a post survey examination of high-resolution color imagery. Deep learning algorithms were trained on color imagery alone, IR imagery alone, IR to color imagery, and fused IR-color imagery.
Results/Conclusions
Preliminary testing has yielded successful detections of ice seals and accurate species classification for ringed and bearded seals. Combining IR and color imagery provided the best performance, reducing overall false positive detections by an order of magnitude. Due to a small image training set for polar bears, models return high false positive rates but still provide successful detections of bears. High variability of IR signature also contributes to this, and ultraviolet imagery has been incorporated to provide a fifth color band for signature stability. Improved image processing allows population estimation analyses to be completed in a timely manner for management considerations. These models will ultimately be applied in-flight, to minimize the number of images collected, reducing the data management, storage, and processing burden.