Color differences between weeds and crops have been used to develop an image-based sensor to detect weeds in wheat and soybean fields for post-emergency herbicide spot spray. Several field factors were found to affect the correct-classification rate (CCR) of the sensor CCR is defined as the ratio between the number of weed-stem pixels correctly classified and the total number of weed-stem pixels in an image. The first factor was soil moisture content. Images of weeds and soil at 10 different moisture contents were analyzed using a statistical method. It was found that, within a large variation range of soil moisture content, the values of relative color indices of soil varied within a sufficiently narrow range, which did not seriously overlap with the range for weed stems. As a result, variations in soil moisture content did not cause significant misclassification between soil and weeds. The second factor was illumination. Images of standard RGB primary-color plates were taken at nine different illumination intensities. The effect of illumination on color indices was found significant. However the variation ranges of the relative color indices on images of the color plates did not overlap, and the variations of the index values followed noticeable patterns, which could be used in color-index calibration. The third factor studied was spatial resolution. Nine different spatial resolutions of weed images were achieved by taking images at different camera-plant distances. The optimum resolution, at which the CCR was maximized, was found to be about 0.5 mm. The classifier was tested on field images obtained under natural lighting conditions. The classifier detected 28-day old yellow foxtail and redroot pigweed with CCRs of 40.2 and 54.9, respectively. The misclassification rate (MCR), which was defined as the ratio between the number of non-weed stem pixels misclassified as weed stem and the total number of non-weed stem pixels, was below 0.2 for both weed species.
展开▼
