Ground-based on-line weed control using computer vision: Analyzing the inference time-accuracy dilemma

Abstract

The detrimental effects of weeds on crop growth and yield present substantial challenges to the agribusiness sector, necessitating the deployment of robust control strategies. The rapid advancement of Computer vision (CV) techniques has driven the integration of ground-based imaging sensors to enable site-specific weed management. The main challenge in weed management revolves around on-line weed detection, which demands a careful balance between inference time and detection accuracy. Finding this balance is very important, as prioritizing a higher number of frames per second (fps) might reduce the detection precision. However, the real-time constraint for on-line weed control remains relatively unexplored. This paper addresses this gap by categorizing proposed approaches based on ground-vehicle configuration and evaluating the real-time requirements for on-line weed control. We comprehensively examine the different components of ground-vehicles including the travel speed, camera settings, and weeding tools to understand the fps required for seamless weed control operation. Results show that for travel speeds below $4km{h}^{-1}$, Deep Neural Networks (DNNs) operating at fps rates lower than 10 Hz are suitable for effective on-line weed detection. However, at higher speeds or with smaller Fields of View, fps demands increase. Our findings further reveal that the relatively relaxed fps requirements of on-line weed control create opportunities for deploying larger DNNs, such as NASNet-A-Large, which can significantly enhance detection accuracy. The operational latency introduced by certain weeding tools further provides additional processing time for DNNs. The continuous advancement of larger DNNs and improvements in hardware offer promising prospects for precise and effective weed management. Future research should leverage these developments, focusing on enhancing detection accuracy rather than optimizing for faster inference times, given the relaxed real-time constraints of ground-based weed control.