2018 ASHS Annual Conference

Improved methods for weed management are considered some of the most pressing needs in organic crop production. Robotic weeding offers a possibility of controlling weeds precisely, particularly for weeds growing near or within crop rows. A study was conducted in 2016 and 2017 at an Iowa State University research farm near Gilbert, Iowa, to examine a machine vision system, based on a Kinect™ V2 sensor, to recognize and localize crop plants at different growth stages, through the fusion of two-dimensional textural data and three-dimensional spatial data, as the first step in constructing an autonomous weeder. Lettuce and broccoli were transplanted at a distance of 61 cm between plants and 91 cm between rows. Weeds at the site consisted of lambsquarter (Chenopodium album), bromegrass (Bromus inermis), pigweed (Amaranthus spp.), waterhemp (Amaranthus rudis), cockspur grass (Echinochloa crus-galli), bindweed (Convolvulus arvensis), purslane (Portulaca oleracea), and white clover (Trifolium repens). Weed populations were not adjusted until after sensor measurements were made. Several feature extraction algorithms were developed for broccoli and lettuce which were heavily infested by the aforementioned weed species. Crop plant recognition algorithms were developed to address the problems of canopy occlusion and leaf damage. With our proposed algorithms, different features in the 3D point cloud data of plants were extracted and used to train plant and background classifiers. For broccoli, the detection rate was 93.1%, and the average localization error was 10.1 mm. For lettuce, the detection rate was 93.7%, and the average localization error was 8.3 mm. The results have shown that 3D-imaging-based plant recognition algorithms are effective and reliable for crop/weed differentiation, which forms the basis for the next steps in developing a fully functioning robotic weeder.