Utilization of Remote Sensing Dataset and a Deep Learning Object Detection Model to Map Siam Weed Infestations

Abstract

Remote sensing data are valuable for detecting, mapping, and managing invasive weed species. This article introduces an innovative algorithm for mapping Siam weed infestations using a bounding box approach from a deep learning object detection model. Supplemented with georeferencing of individual aerial images obtained from an uncrewed aerial vehicle (UAV), the study demonstrates the potential for large-scale mapping. High-resolution RGB images were used to develop a Siam weed detection model with YOLOv5, achieving 0.95 precision, 0.82 recall, and 0.88 F1-Score, indicating suitability for detecting flowering Siam weed. The validated model was applied to an independent single-flight dataset, enabling consecutive detections on high-resolution images. Georeferencing the detections was accomplished using a customized raycasting algorithm. To address fragmented and duplicated detections from overlapping images, the authors proposed buffering, dissolving, and filtering the georeferenced detection boxes. Therefore, revealing an extensive Siam weed infestation boundaries and pattern for a whole flight area. The developed algorithm successfully maps Siam weed infestations using an object detection bounding boxes approach on individual aerial images, which is potentially advantageous for speed and scalability. It presents an opportunity to upscale weed detection and mapping with reduced overlapping percentage and higher payload aircraft, such as helicopters, allowing for larger-scale surveys and enhancing weed management teams' capacity to monitor extensive landscapes.

Keywords

• Automated mapping
• georeferencing
• invasive species
• uncrewed aerial vehicles (UAVs)
• YOLOv5