Optimisation of weed management by image segmentation in precision agriculture

Abstract

Accurate weed detection remains crucial for ultra-localised control in robotic solutions, simulating manual weeding in agriculture. Although many studies have been conducted in the field of weed detection using machine learning, most have focused mainly on direct detection, which can present challenges in the face of weed diversity. In this study, we propose an integrated approach based on vegetation/soil segmentation, followed by discrimination between crops and weeds using an object detector. Segmentation models such as UNet, FPN, and LinkNet have been thoroughly trained to discriminate efficiently between vegetation and soil. The results obtained are promising, with the trained models being able to generate binary images (masks) with an accuracy (Jaccard and Dice similarity indices) of over 89%. In addition, the execution speed reached 217 frames per second (fps). The integration of the localisation results from the detection model with the segmented images provides a robust method for accurately determining the position of weeds in the agricultural context, opening up new prospects for automated, targeted weed control solutions.