Robotic weed control system for tomatoes

Abstract

A real-time intelligent robotic weed control system was developed for selective herbicide application to in-row weeds using machine vision and precision chemical application. The image processing algorithm took 0.34s to process one image containing 10 plant objects, representing a 11.43 cm by 10.16 cm region of seedline, and allowing the prototype robotic weed control system to travel at a continuous rate of 1.20 km/h. In actual field trials conducted in a commercial processing tomato field, the robotic weed control system correctly identified and didn't spray 75.8% of the tomato plants and correctly sprayed 47.6% of the weeds. The color segmentation look-up-tables made in hue, saturation, intensity color space were generally better than those made in normalized red, green, blue and un-normalized red, green, blue color spaces. Overall, look-up-tables built only with hue gave the best performance, correctly classifying 77.8% of color pixels. In validation tests with 290 field images from 13 different commercial processing tomato fields, the image processing algorithm correctly identified 58.5%-80.7% of tomato cotyledons, 9.0%-21.2% of tomato true leaves, 8.9%-12.9% of tomato leaves that were curled, occluded, bug-eaten or partially hidden by the edge of the image, and 93.0%-95.5% of weeds using plant area, length to perimeter ratio, and occupational ratio. For separation of occluded plant leaves, 5 modifications to the watershed algorithm were investigated. The performance of opening, feature criteria, and combined opening and feature criteria modifications were better than others. The recognition of occluded tomato cotyledons and true leaves improved 33.3% and 41.1%, respectively, after the modified watershed algorithm was applied to occluded objects. The angle of tomato cotyledon orientation for some varieties changed at different times of day. The critical orientation angle for tomato cotyledon recognition was estimated as 27.5$\sp\circ$ with the vertical axis of the plant. Spray trials with 2.54 cm and 1.27 cm diameter targets showed that the robotic spraying system correctly sprayed these simulated "weeds", targeting their centers within an average spatial error of 0.51-1.36 cm and a standard deviation of 0.21-.71 cm on three different ground surfaces.