Development of full-scale machine vision scheme of multi-camera systems on boom sprayer for real-time management of Colorado Potato Beetles

Abstract

Precision agriculture demands site-specific pest management to reduce chemical inputs and environmental harm, yet real-time detection of small insect pests from mobile platforms re-mains challenging. The Colorado potato beetle (CPB), measuring less than 10 mm in body length, is among the most economically destructive pests in potato production globally and has developed resistance across all major insecticide classes. This research addresses the gap between laboratory-validated detection algorithms and field-deployable systems by developing, integrating, and evaluating a machine vision scheme for real-time CPB detection under commercial operating conditions. A field dataset was constructed from imagery captured using static handheld and moving sprayer-mounted cameras across two growing seasons, with bounding-box annotation under variable lighting, diverse canopy structures, and natural backgrounds. An image-cropping-based pre-processing technique was developed to preserve fine-scale morphological features suppressed by conventional input resizing, directly addressing the tiny object detection challenge in high-resolution agricultural imagery. Three deep learning architecture — YOLOv5, YOLOv7, and Faster R-CNN — were trained and benchmarked; YOLOv5 was selected for deployment on its favorable accuracy–latency trade-off, achieving 78 % accuracy on independent test data while sustaining real-time throughput. A modular software pipeline was engineered for concurrent multi-camera stream processing, combining TensorRT-accelerated inference with deterministic spray command generation and network-based communication with sprayer control systems. End-to-end latency measured approximately 120 milliseconds across four simultaneous camera feeds. Full-scale machine vision scheme deployment was implemented on a Case IH Patriot 3240 boom sprayer with eight machine vision systems across a 30-meter boom, each handling three to four downward-facing cameras. Engineering challenges resolved included distributed power supply, environmental enclosure de-sign, boom folding compatibility, and synchronized multi-unit operation under field vibration. Field validation at a commercial spraying speed of 2.7 m/s confirmed stable real-time operation, with detection accuracy exceeding 80 %, precision near 90 %, and recall approaching 90 %. Furthermore, a novel five-crop dataset expansion technique increased effective training data fivefold while preserving object scale and introducing natural background diversity; benchmarking with state-of-the-art YOLOv11 architecture demonstrated mean average precision exceeding 85 % at standard IoU thresholds. This research establishes a validated framework for insect-specific precision spraying, with contributions in tiny object detection, real-time multi-camera processing, and mobile agricultural system integration.