WeedCAM: An edge-computing camera system for multi-species weed detection in sugar beet production fields

This study introduces WeedCAM, a low-cost, near real-time, edge-computing camera system for multi-species weed detection, built on a Raspberry Pi 5 and integrated with a GPS module and LoRa board for geolocation and data transmission. A three-phase framework, including data acquisition, model fine-tuning, and deployment, is proposed to implement detection models on WeedCAM. A total of 5734 high-resolution (4K) images were automatically collected using WeedCAM, producing a dataset with a long-tail distribution that poses challenges for model training. To address this, Repeat Factor Sampling and Focal Loss were applied during the fine-tuning. Seven object detection models were evaluated, including YOLOX-S, YOLOX-L, Faster R-CNN, Cascade R-CNN, Deformable-DETR, and DINO. Finally, WeedCAMs with embedded trained models were deployed in the field on pivot-mounted and ground-based installations, detecting weeds at 30-min intervals and transmitting results to a customized gateway via LoRa. The gateway parsed and mapped these results to our custom-designed website for visualization. Our best model, DINO-Swin/L, set the performance benchmark with a 76.0 overall mAP (IoU = 0.5) and strong per-species scores for kochia (76.4 mAP), Palmer amaranth (77.3 mAP), and volunteer corn (75.9 mAP) at 4K resolution image. Despite this, YOLOX-L was deployed on the WeedCAM, as its efficient 8-min processing cycle represented the better trade-off between accuracy and speed. Field evaluation confirmed that WeedCAM effectively identified weed species and quantities under varying lighting conditions, camera angles, and soil moisture levels during irrigation events. These results demonstrate the practicality of deploying WeedCAM edge-computing deep learning systems for near real-time, multi-species weed detection under sugar beet fields. • A low-cost edge-computing camera system was proposed for near real-time, multi-species weed mapping. • Pivot-mounted camera system enabled continuous, autonomous season-long field monitoring. • A three-phase pipeline for robust WeedCAM deployment was proposed. • Two datasets aligned with critical herbicide application windows were created. • Field tests validated the generation of accurate multi-species weed pressure maps.