Abstract
As the world's population has grown and mobility has become more accessible, human activities have expanded into remote areas and deserts. These activities (typically desert safaris) generate much left-behind waste. Untreated left-out wastes contribute directly to gas emissions that exacerbate global warming, especially litter such as plastics, cans, or any other resistant material. We present a drone-based solution to find waste in extra-urban environments (mainly deserts). In this paper, we use an autonomous unmanned aerial vehicle (drone) that flies over a desert area to capture a video stream to locate and identify waste objects such as plastic bottles. Our solution is still a work in progress; hence, in this paper, we have explored an object detection algorithm using transfer learning (few-shot-learning), where a pre-trained classifier (YOLO) is retrained with our relatively small dataset DroneTrashNet. Our goal is to benchmark existing models and approaches for object detection and classification in order to provide a baseline for selecting the most appropriate models for detection and classification tasks in extra-urban contexts. Preliminary results show an achieved mAP of 42% in specific conditions, representing a significant improvement over base models.
Introduction
As global tourism and events—such as desert safaris or international sporting competitions—expand into remote regions, the environmental footprint of man-made waste grows proportionally. In arid environments, litter like plastic bottles and cans contributes to gas emissions and long-term ecological degradation. Monitoring these vast, inhospitable areas is difficult for human crews. While computer vision has advanced urban waste management, desert environments pose unique hurdles: homogenous sandy backgrounds, harsh reflective lighting, and the tendency for wind to partially bury objects, leaving only fragments visible. Existing "off-the-shelf" models trained on standard datasets often fail to recognize these obscured, sun-bleached objects against a desert backdrop.
System Architecture and Methodology
The proposed solution utilizes an autonomous Unmanned Aerial Vehicle (UAV) to capture high-definition video streams while flying at low altitudes. This data is transmitted to a cloud-based data center for processing and inference.
The core of the detection system is the YOLOv4 (You Only Look Once) architecture, selected for its balance of high accuracy and inference speed (up to 45 FPS). Because the custom dataset was relatively small, the team employed transfer learning. A YOLOv4 model pre-trained on the COCO dataset was fine-tuned using desert-specific images, allowing the model to retain a general understanding of shapes while learning the specific visual signatures of trash in the sand.
The DroneTrashNet Dataset
To bridge the gap between urban and extra-urban waste detection, the researchers developed DroneTrashNet. This specialized dataset consists of annotated video captured by a drone at low altitude. The dataset was augmented from 200 to 400 frames via 90-degree rotations to improve model robustness.
Crucially, to simulate real-world conditions, approximately 50% of the objects were manually half-buried in the sand. This "occlusion" training prepares the AI for the reality of environmental monitoring where objects are rarely fully visible. The dataset includes six labels: Glass, Plastic, Bottle, Cardboard, Metal, and a generic "Waste" tag.
Results and Discussion
The specialized YOLOv4 model achieved a mean Average Precision (mAP) of 42%. This is a substantial improvement over the standard YOLOv4 model (trained only on COCO), which achieved a meager 13% mAP when applied to desert waste. This "context gap" underscores the necessity of domain-specific retraining for specialized environments.
Qualitative results demonstrate that the model can successfully identify and bound bottles and generic waste even in high-glare, sandy conditions. The 42% mAP is comparable to the original YOLOv4 performance on the COCO dataset (~44%), suggesting the specialized model is performing near its architectural ceiling despite the limited training data.
Conclusion and Future Work
While these preliminary results are promising, several challenges remain. The current reliance on cloud computing introduces latency; future iterations will explore lighter architectures like YOLOv6 or Fast R-CNN to enable edge computing directly on the drone hardware. Additionally, the team plans to expand the DroneTrashNet dataset with images captured at varying altitudes (5m to 20m) and camera angles to improve generalization. The ultimate goal is a fully integrated system that automatically alerts local garbage collection services with precise GPS coordinates of detected litter.
Summary updated