Improved UAV Opium Poppy Detection Using an Updated YOLOv3 Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Illicit opium poppy cultivation poses a significant challenge for drug crime prevention.
• Current detection methods using unmanned aerial vehicle (UAV) imagery are often slow due to manual interpretation.
• There is a need for automated and efficient methods for identifying poppy plants from aerial data.

Purpose of the Study:

• To develop and evaluate an improved You Only Look Once version 3 (YOLOv3) model for rapid and accurate detection of opium poppy plants in UAV imagery.
• To assess the impact of different backbone networks on detection performance and speed.
• To optimize the model by incorporating a Spatial Pyramid Pooling (SPP) unit and Generalized Intersection over Union (GIoU) loss.

Main Methods:

• Utilized the YOLOv3 network architecture with various backbone networks, selecting MobileNetv2 (MN) as optimal.
• Integrated a Spatial Pyramid Pooling (SPP) unit to enhance feature extraction.
• Employed Generalized Intersection over Union (GIoU) for coordinate loss calculation.
• Applied a sliding window method for processing complete UAV images.

Main Results:

• The SPP-GIoU-YOLOv3-MN model achieved a high average precision of 96.37%.
• Detection speed reached 29 frames per second (FPS) on an RTX 2080Ti platform.
• Processing complete UAV images took approximately 2.2 seconds per image, a tenfold increase in speed compared to visual interpretation.
• The model demonstrated a 1.62% improvement in average precision over baseline methods.

Conclusions:

• The developed SPP-GIoU-YOLOv3-MN model offers a significant advancement in the efficiency and accuracy of detecting illicit opium poppy cultivation from UAV imagery.
• This automated approach is highly suitable for rapid identification in diverse environments like residential areas and farmland.
• The method enables timely intervention and supports law enforcement in combating drug-related crimes effectively.