Local and Landscape Scale Detection of Reindeer Carrion in Svalbard Using High-Resolution Imagery

By Julia Engblom

Abstract
Decomposing organic matter is crucial for biodiversity and ecosystem stability, especially in rugged and nutrient-poor terrains like those of Svalbard, as it returns the energy and nutrients accumulated in living organisms back into the biosphere. Despite the ecological impact of dead animal biomass, or carrion, studies on their decomposition are scarce and lack a universal framework making them largely overlooked in ecosystem models. There is a considerable gap in carrion production and quantifying it is a vital first step in determining their function in nutrient cycles.

Remote sensing combined with computer vision has become a valuable tool for wildlife detection due to its efficiency in time, workload, and large-scale coverage. To explore to what extent reindeer carrion can be automatically detected in the Arctic tundra of Svalbard, three convolutional neural networks (CNNs) were developed using very high-resolution imagery. This paper presents the results of using images from an uncrewed aerial vehicle (UAV) and the WorldView-3 (WV-3) satellite to compare the extent of detection at local and landscape scale.

The two object detection models, You Only Look Once (YOLO), outperformed the semantic segmentation model, U-Net, although the U-Net performance likely suffered from an error during metrics computation. YOLOv7 successfully detected most of the carrion in both the UAV and WV-3 testing datasets. The YOLOv7 UAV model achieved state-of-the-art performance, with a precision of 96.8%, a recall of 76.4%, and a mean average precision (mAP0.5) of 88.2%, while the YOLOv7 WV-3 model requires further refinement, as it reached a precision of 45.4%, a recall of 45.5%, and an mAP0.5 of 23.9%. The performance of all models was limited by the availability of training data, particularly the WV-3 models, which were trained on just 45 carrion images. Yet, the YOLOv7 model's ability to identify most carrion despite the small dataset is a very promising result for future landscape level carrion detection. In conclusion, while carrion detection at local scale outperformed landscape scale, substantially increasing the number of training images, especially satellite, will likely improve the performance of both models, making automatic detection and accurate quantification of reindeer carrion production feasible in the future.