Design and development of a low-cost industrial prototype of spot-specific spraying system for potato weed detection using deep learning

Conventional sprayers uniformly distribute agrochemicals on crops, even though weed distribution is typically patchy, leading to increased costs and environmental pollution. Spot-specific AI-driven sprayer offers a solution to optimize this challenge. The primary objective of this research is to develop a low-cost, smart, and fast-response industrial prototype of the spot-specific spraying system integrated with deep learning-based object detection techniques. The study aims to distinguish various weed species to enable the precise application of herbicides on potato crops. To ensure a real-world representation, a diverse dataset was collected from the real potato fields, covering various climatic conditions, irrigation methods, and timestamps. This data was utilized for training YOLOv8, YOLOv9, and YOLOv12 detection models, resulting in detection accuracies of 90%, 92%, and 93% respectively. The results indicate that the YOLOv12 model obtained higher detection accuracy while localizing the weed instances accurately. Additionally, the average processing speed of 35 frames per second and 320 ms lag time between image acquisition and activation demonstrate the efficacy of the adopted YOLOv12 model. Following the result analysis, the high detection model (YOLOv12) is integrated with the spot-specific sprayer system for real-time detection of weeds in potato fields. The system accurately identified the weeds in the potato field and triggered the nozzles accordingly. The microcontroller is responsible for controlling this system to ensure the precise application of weedicide. The results indicate that the developed spot-specific sprayer system not only achieves a nearly 40% reduction in herbicide usage, leading to cost savings, but also reduces the environmental impact.