Camera-based estimation of sugar beet stem points and weed cover using convolutional neural networks

This study demonstrates a camera system that can determine the location of sugar beets automatically by analyzing images. The system works with high weed densities, and it can detect sugar beets that take up less space than the weeds, and also sugar beets that are partly occluded by weeds. The system processes top-down RGB images of sugar beets to both determine plant cover and stem locations. Two convolutional neural networks were trained to perform pixel-wise classification of sugar beets and weeds with sub-centimeter precision and to detect stem points of sugar beets. When classifying sugar beets and weeds, an intersection over union of 0.78 and 0.57 for sugar beets and weeds, respectively, was obtained. Predictions can be used to estimate weed cover and to determine weed competition and the need for weed removal. Furthermore, it can be used for mapping growth of the sugar beets throughout the field. Among 222 test sugar beets, the stem locations of 93.3% were detected with an average error of 11.9 mm between annotated and detected points.