Counting animals in aerial images with a density map estimation model

Animal abundance estimation is increasingly based on drone or aerial survey photography. Manual post-processing has been used extensively, however, volumes of such data are increasing, necessitating some level of automation, either for complete counting or as a labour-saving tool. Any automated processing can be challenging when using such tools on species that nest in close formation such as Pygoscelis penguins. We present here a customized CNN-based density map estimation method for counting of penguins from low-resolution aerial photography. Our model, an indirect regression algorithm, performed significantly better in terms of counting accuracy than standard detection algorithm (Faster RCNN) when counting small objects from low-resolution images and gave an error rate of only 0.8 percent. Density map estimation methods as demonstrated here can vastly improve our ability to count animals in tight aggregations, and demonstrably improve monitoring efforts from aerial imagery.Funding provided by: World Wildlife FundCrossref Funder Registry ID: http://dx.doi.org/10.13039/100001399Award Number: B095701The British Antarctic Survey currently holds an archive of colour digital aerial photography from the Antarctic Peninsula and South Shetland Islands acquired between November and December 2013, and partially re-flown in November 2015. The archive contains images from approximately 140 Pygoscelis penguin colonies selected for a range of species, population sizes and topographic settings. The images were acquired using a large-format Intergraph DMC mapping camera, with a resolution of about 12 cm or better. The images each have a footprint of about 1600 m * 1000 m and were flown with 60% overlap to allow stereo-cover. For the images to be useful as part of an automated penguin counting process they needed significant pre-processing to geolocate them and remove terrain distortions inherent to the perspective view of a camera image. This processing comprised: 1) the stereo-images were used to extract a Digital Elevation Model (DEM); 2) the images were ortho-rectified to the DEM to remove terrain effects; 3) the processed images were mosaicked; and then, 4) cut into standard-sized (448 * 448 pixels) tiles for counting. This process ensures that the images are accurately located and scaled to enable accurate ground area measurements and hence penguin density estimates. Without the DEM and orth-rectification pre-processing, the counts would not have a reliable ground area estimate. Stages 3) and 4) also ensure that each penguin only appears once in the dataset. The process to create the DEM is relatively complex, and utilized BAE Systems Socet GXP photogrammetry software to generate DEMs, ortho-rectify the images and prepare geo-referenced mosaics for each colony. Aerial imagery from the Intergraph DMC mapping camera allowed multiple penguin colonies to be photographed within a single survey flight on board a deHavilland Twin Otter. This is advantageous when synoptically surveying large areas of terrain where many penguin colonies may occur. Pygoscelis penguins generally breed within colonies that comprise a single species, although on occasions there may be two species in close proximity where their colony boundaries interdigitate. Our study did not use colonies where two species co-occur as we only considered separate colonies of gentoo (Pygoscelis papua), Adélie (P. adeliae), and chinstrap (P. antarctica) penguins