Automized segmentation of animal paths in UAS Imagery using deep learning - a case study from the Kruger National Park

The behavior of terrestrial animals gives insights into their decision-making in relation to their local or inter-regional environment. Governed by this interaction, animals move across the landscape, leaving traces in the form of broad paths, trodden down trails, and stopping and resting areas. While GPS and radio-tracking provide valuable data on individual animal movements they are spatially and temporally limited due to fix errors and battery life. Hence, being able to identify continuous animal path systems is of great importance for understanding animal behavior and habitat utilization, complementing discrete geolocation data. In this work, we used photogrammetry-derived Red-Green-Blue (RGB) orthomosaics and digital surface models from Unoccupied Aerial System (UAS) imagery obtained in the Kruger National Park, South Africa. To automatically map animal paths, we used a UNet, a Deep Learning architecture. To generate training data, we developed a semi-automatic labeling approach to estimate the width of manually labeled animal paths, published alongside this work as an R package called checkRtrack. We investigated how different methods to derive animal path training data from imagery can enhance our network path segmentation performance. We show that our network is capable of correctly predicting up to 83% of animal paths on additionally created test data. High precision in path width estimation for training data generation was found to strongly improve the results of path segmentation by our network. • The automatized detection of animal paths through a deep convolutional neural network. • Semi-automatic DSM-based estimated path width using the checkRtrack R-package. • Investigation on annotation subjectivity regarding its effects on CNN performance. • Offers a spatially scalable approach for habitat assessment analysis using UAS data.