Towards machine vision based site-specific weed management in cereals

Abstract

Arable weeds are generally distributed in patches, while herbicides are applied uniformly. Herbicides can be saved if only the patches are sprayed, i.e. patch spraying (PS). Bottlenecks for cost-effective PS are weed monitoring technology and valid technology-based decision rules for PS (thresholds). The novel machine vision algorithm Weedcer has been developed as an efficient weed monitoring tool for PS. Weedcer estimates the proportions of young weed leaves and cereal leaves in high resolution red–green–blue images. We conducted field trials to test relative weed cover (RWC) and relative mayweed cover (RMC) estimated by Weedcer as decision rules for PS . RWC is the total weed cover divided by the total plant cover and RMC is the mayweed cover divided by the total plant cover. The main criterion for evaluation and basis of these thresholds was the measured grain yield. Images (about 0.06-m 2) were acquired with a GPS guided autonomous field robot in spring, the normal time for spraying seed-propagated broadleaf weeds in both winter – and spring cereals in Norway. Three map-based trials (weed monitoring and spraying in two separate operations) showed that mean RWC per management unit (12.0 × 12.5-m) was generally adequate. In winter wheat heavily infested with scentless mayweed ( Tripleurospermum inodorum (L.) Sch.Bip.) and/or scented mayweed ( Matricaria recutita L.), the mean RMC per management unit was more adequate. Progress during the project allowed three additional trials conducted in real-time (weed monitoring and spraying in the same operation). These were conducted with the robot in spring cereals, and showed that a weighted moving average of RWC per image was adequate. The sprayed and unsprayed management units in these trials were minimum 3.0 × 3.0-m and 0.5 × 3.0-m, respectively. Results indicated that the Weedcer-based thresholds should be lower in wheat ( Triticum aestivum) than in barley ( Hordeum vulgare).