Weed Detection Using Color Machine Vision

Many weed species have reddish stems, but stems of wheat and soybean are green. These color features wereused in this study to establish a simple weed-detection method using a color machine-vision system. This method is morepractical than texture- or shape-based methods because of its low sensitivity to canopy overlap, leaf orientation, camerafocusing, and wind effect. Four types of relative color indices formed by RGB gray levels were designed. The mosteffective combinations of these color indices were selected using a statistical method. These combinations were used asthe input variables for a statistical classifier based on discriminant analysis (DA) and two artificial neural-network (NN)classifiers. These classifiers were trained and tested using three weed species (Johnsongrass, redroot pigweed, and yellowfoxtail) with soybean and three weed species (wild buckwheat, cheat, and field bindweed) with wheat. Preprocessing andpostprocessing algorithms were developed to shorten the processing time and to reduce noise. The results showed that thestatistical DA classifier was more accurate than the NN classifiers in classification accuracy. The least-square means ofthe classification rates using the DA classifiers for soybean and wheat were 54.9% and 62.2%, respectively. Themisclassification rates for most weed species were below 3%. Because the reddish colors on the stems of some weedspecies vary as the plants grow, an in-field calibration procedure will be needed to make the classifiers more adaptive todifferent circumstances.