Detection of Endangered Deer Species Using UAV Imagery: A Comparative Study Between Efficient Deep Learning Approaches
Agustin Roca, Gastón Castro, Gabriel Torre, Leonardo J. Colombo, Ignacio Mas, Javier Pereira, Juan I. Giribet · 2025
Read Detection of Endangered Deer Species Using UAV Imagery: A Comparative Study Between Efficient Deep Learning Approaches by Agustin Roca, Gastón Castro, Gabriel Torre, Leonardo J. Colombo, Ignacio Mas, Javier Pereira, Juan I. Giribet (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-00154.txt.
Introduction
Agustín Roca and co-authors teach drones to spot endangered marsh deer hiding in tall vegetation. Their 2025 paper compares recent detection networks, including variants of YOLOv11 and RT-DETR, on aerial images where a deer fills only a few pixels and plants cover the rest. Finding animals this small and hidden is the core difficulty.
The team adds hand-drawn outline masks to their datasets, which lets one model train with a segmentation head that traces each animal's shape. That shape-aware model wins the comparison, finding more deer with fewer misses than its rivals. The authors present the pipeline as a scalable tool for wildlife monitoring, where accurate counts guide how rangers protect a threatened species.
Catalog record
| CREATOR | Agustin Roca, Gastón Castro, Gabriel Torre, Leonardo J. Colombo, Ignacio Mas, Javier Pereira, Juan I. Giribet |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-00154
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Detection of Endangered Deer Species Using UAV Imagery: A Comparative Study Between Efficient Deep Learning Approaches Agustin Roca, Gastón Castro, Gabriel Torre, Leonardo J. Colombo, Ignacio Mas, Javier Pereira, Juan I. Giribet Abstract This study compares the performance of state-of-the-art neural networks including variants of the YOLOv11 and RT-DETR models for detecting marsh deer in UAV imagery, in scenarios where specimens occupy a very small portion of the image and are occluded by vegetation. We extend previous analysis adding precise segmentation masks for our datasets enabling a fine-grained training of a YOLO model with a segmentation head included. Experimental results show the effectiveness of incorporating the segmentation head achieving superior detection performance. This work contributes valuable insights for improving UAV-based wildlife monitoring and conservation strategies through scalable and accurate AI-driven detection systems. ©2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. I Introduction Monitoring wildlife is crucial for understanding and preserving ecosystems. Traditional observation methods, however, are often labor-intensive, costly, and constrained in coverage [ 1 , 2 ] . The integration of Unmanned Aerial Vehicles (UAVs) offers a transformative approach, enabling extensive area surveillance and remote data collection. Despite these advantages, challenges persist in effectively detecting and identifying species, particularly those that are small or well-camouflaged, such as deer, in aerial imagery. This work examines the application of artificial intelligence techniques to automate the identification of deer in aerial images, addressing a significant challenge in wildlife conservation. Recent efforts have focused on leveraging UAV-based monitoring for the protection of endangered deer species in diverse ecological settings. These initiatives involve multidisciplinary teams comprising scientists, engineers, and conservationists, working towards improving population monitoring and habitat assessment. A key challenge in these efforts is the automatic detection of deer using aerial imagery, which demands robust image processing and machine learning techniques. UAV surveys have been conducted in various regions, collecting high-resolution imagery to support conservation strategies. The implementation of AI-based detection systems is a crucial step in automating the identification process, reducing human effort, and enhancing monitoring accuracy. One such initiative is the Pantano Project [ 3 ] , dedicated to the conservation of marsh deer ( Blastocerus dichotomus ) in the Paraná Delta, Buenos Aires, Argentina. This project, involving scientists from research centers and universities, aims to study the ecology and behavior of marsh deer and develop effective conservation strategies. UAVs have played a critical role in this initiative by enabling remote detection of marsh deer in their natural habitat. Aerial surveys conducted as part of the project have facilitated extensive data collection, which has been instrumental in training AI-based detection models. Collaboration with researchers and trained volunteers has further enhanced the effectiveness of these efforts. This study presents the initial results of developing an automated detection system for deer species using deep learning algorithms. Utilizing datasets generated from aerial surveys, object detection models based on YOLO (You Only Look Once) [ 4 , 5 ] and RT-DETR (Real Time DEtection TRansformer) [ 6 ] have been trained and evaluated. Some detections using these trained models can be seen in Figure 1 . The system’s performance is assessed by comparing its detections against expert-verified census data, providing insights into its reliability and potential for large-scale deployment. Fig. 1 : Visual comparison of predictions for a sample image from the test set. It shows the predicted class and the confidence score of each detection II Related Work Over the past decade, UAVs have gained traction as a tool for wildlife observation. Linchant et al. [ 7 ] provide an extensive review of advancements and challenges in this field up to 2015. Their study highlights key technical limitations, including restricted coverage of large geographic areas, emphasizing the need for improved sampling protocols, inventory methods, and statistical analyses. Additionally, they underscore the potential of combining thermal sensors with high-resolution optical imaging to enhance species differentiation. Another major challenge identified is the lack of regulatory frameworks supporting UAV-based wildlife monitoring. Since 2015, substantial progress has been made in both regulatory aspects—albeit at a slower pace—and in refining methodologies for data collection and analysis [ 8 , 9 ] . Advances in sensor fusion techniques have further improved species detection, while artificial intelligence-driven automatic detection strategies have emerged as valuable tools for streamlining species identification and classification. A variety of approaches have been explored for detecting wildlife in aerial imagery, including object detection algorithms, semantic segmentation techniques, and deep learning-based methods [ 10 , 11 , 12 ] . For instance, Barbedo et al. [ 13 ] leveraged Convolutional Neural Networks (CNNs) to monitor cattle herds using UAV-acquired imagery. In another study, Brown et al. [ 14 ] utilized high-resolution UAV images to establish precise ground-truth data, implementing YOLOv5-based object detection for livestock localization and counting in Australian farms. Deep learning models, particularly YOLO, have demonstrated significant potential for UAV-based wildlife detection tasks. The effectiveness of deep learning in UAV wildlife monitoring is highly dependent on access to large-scale datasets for training algorithms [ 15 ] . However, the scarcity of extensive aerial wildlife image datasets has posed a challenge, often requiring adaptation of object detection models originally designed for natural scene images. This adaptation, though widely used, is not always optimal for wildlife detection applications [ 16 ] . Consequently, UAV-based detection algorithms continue to struggle with issues such as low accuracy, limited robustness, and suboptimal real-world performance [ 17 , 18 ] . Several studies have sought to address these limitations. For example, Mou et al. [ 19 ] introduced a real-time species detection model based on YOLOv7, optimized for detecting small targets in UAV-based monitoring. Their work also presented the Wildlife Aerial Images from Drone (WAID) dataset, consisting of 14,375 high-quality UAV images across multiple environmental conditions, featuring six different wildlife species and diverse habitat types. Comparative experiments demonstrated that the SE-YOLO model outperforms others, achieving a mean average precision ( mAP ) of 0.983, with a particularly high mAP of 0.926 for animals occupying fewer than 50 pixels in an image. Similarly, Lyu et al. [ 20 ] explored RGB and thermal image fusion for deer detection, showing that transferring a model trained on visible spectrum images to thermal imaging is challenging due to the limited detail in thermal data. However, with an appropriately designed neural network, detection performance significantly improves. Fig. 2 : Images with different animals spotted during UAV test flights, including capybaras, deer, cows, and birds. In a previous related study [ 21 ] , the present authors proposed an approach for the automatic detection of endangered deer species using UAV aerial images combined with the YOLO framework. The model, trained on high-resolution RGB images manually annotated (the same dataset used in this work), achieved an mAP@10 of 0.950 in real-world conservation applications. However, when tested with the Pampas deer, a different cervid species, the model exhibited a significant performance drop, highlighting the need for species-specific training data. The current