Weed detection using color machine vision

Weed detection algorithms were developed using RGB color images. The emphasis was on weed species associated with Kansas winter wheat and soybean. The study was completed through five phases.; The first phase looked into the effect of soil moisture content on weed detection. It was found that, by using relative color indices, separation between weed stem and soil would not be seriously affected by variations in soil moisture content.; The second phase investigated the effects of illumination on color indices of primary color plates. The effects were found insignificant. Changes of color indices with illumination followed patterns which could be used in color-index calibration.; The third phase determined the optimum spatial resolution for best classification. The optimum horizontal and vertical resolution were found to be 0.052cm/pixel and 0.042cm/pixel, respectively.; During the fourth phase, algorithms were developed for weed stem segmentation and noise removal using images of six leading weed species commonly found in wheat and soybean fields. Three detection methods employing statistical and neural-network techniques were developed. All detection methods gave satisfactory results with the statistical classifier outperformed the neural-network classifiers.; In the last phase, the algorithms were tested on field images obtained under uncontrolled natural lighting conditions. The outcome proved that the detection algorithms were capable of detecting weeds in soybean and wheat fields with significantly low misclassification.