It appears you don't have support to open PDFs in this web browser. To view this file, Open with your PDF reader
Abstract
Sub-Saharan Africa is a hotbed of remarkable terrestrial biodiversity, home to a unique diversity of mammals. Unfortunately, this richness is threatened by the growing impact of human activities. While wildlife populations are declining, livestock numbers have been increasing for decades. With significant population growth predicted for the region this century, the pressures on biodiversity are likely to intensify. It is therefore imperative to closely monitor wild and domestic mammal populations. The conventional method of aerial counting using systematic sampling is still widely used to census these populations in open areas. However, the use of on-board photographic sensors on various remote sensing platforms offers the potential to improve and standardize traditional methods. However, processing the large quantities of data generated by these sensors remains a major challenge. In this context, the use of automatic approaches based on deep learning, a branch of artificial intelligence, appears to be a promising solution. The objective of this thesis is therefore to evaluate the effectiveness of the combined use of remote sensing and deep learning for the multi-species census of large African mammals. The research spans several protected areas and considers various mammal species, both wild and domestic.
Firstly, I assessed the potential of pre-existing convolutional neural network architectures to automate the detection and identification of wild species in ultra-high resolution (UHR) images (Chapter 2). Three architectures were tested on a dataset comprising six large mammal species. The best model, achieving a mean Average Precision (mAP) of 80%, was applied to an independent dataset from Garamba National Park, Democratic Republic of Congo. It showed superior detection performance to previous studies in similar habitats, opening up promising prospects. However, detection limits were observed for the smallest species (warthog, Phacochoerus africanus), and a drop in precision was observed in herd situations for African elephant (Loxodonta africana) and buffalo (Syncerus caffer), due to a high number of false positives.
Secondly, I developed a novel deep learning architecture named HerdNet in response to the limitations of pre-existing models (Chapter 3). HerdNet is a point-based object detector inspired by crowd-counting techniques. It has been tested on oblique images of domestic herds of camel (Camelus dromedarius), donkey (Equus asinus), sheep (Ovis aries) and goat (Capra hircus)from the Ennedi Natural and Cultural Reserve in Chad. HerdNet demonstrated better detection and counting accuracy than previous methods, on both oblique (+26% of F1 score) and nadir UHR images (+32%). In addition, it solves the problem of false positives in dense herd situations, with proximity-invariant precision. Although species identification could be improved, the practical benefits and potential use of HerdNet were discussed, promising a significant reduction in the human interpretation time associated with aerial surveys.
Thirdly, I evaluated the contribution of oblique UHR imagery and deep learning on systematic aerial sample surveys, in a semi-automatic detection context (Chapter 4). I first quantified the reduction in human workload associated with the manual interpretation of oblique images acquired by an on-board camera system on light aircraft. HerdNet was used to detect, count and identify 12 animal species in the Queen Elizabeth Protected Area, Uganda, resulting in a 74% reduction in the number of images to be interpreted by humans.





