Recent computer vision techniques based on convolutional neural networks (CNNs) are considered state-of-the-art tools in weed mapping. However, their performance has been shown to be sensitive to image quality degradation. Variation in lighting conditions adds another level of complexity to weed mapping. We focus on determining the influence of image quality and light consistency on the performance of CNNs in weed mapping by simulating the image formation pipeline. Faster Region-based CNN (R-CNN) and Mask R-CNN were used as CNN examples for object detection and instance segmentation, respectively, while semantic segmentation was represented by Deeplab-v3. The degradations simulated in this study included resolution reduction, overexposure, Gaussian blur, motion blur, and noise. The results showed that the CNN performance was most impacted by resolution, regardless of plant size. When the training and testing images had the same quality, Faster R-CNN and Mask R-CNN were moderately tolerant to low levels of overexposure, Gaussian blur, motion blur, and noise. Deeplab-v3, on the other hand, tolerated overexposure, motion blur, and noise at all tested levels. In most cases, quality inconsistency between the training and testing images reduced CNN performance. However, CNN models trained on low-quality images were more tolerant against quality inconsistency than those trained by high-quality images. Light inconsistency also reduced CNN performance. Increasing the diversity of lighting conditions in the training images may alleviate the performance reduction but does not provide the same benefit from the number increase of images with the same lighting condition. These results provide insights into the impact of image quality and light consistency on CNN performance. The quality threshold established in this study can be used to guide the selection of camera parameters in future weed mapping applications.

Weed competition is one of the most limiting factors affecting crop yield and profitability. Robotic weeding systems have demonstrated their potential to save herbicide usage and thereby minimize costs and adverse impacts on the environment. We introduce the software and hardware design of an automatic system for micro-volume herbicide spray using a mobile robot for early-stage weed control. The system is equipped with a stereo camera, one inertial measurement unit, and multiple linearly actuating spray nozzles. To enable the system, we propose a new scene representation from the perspective of spray operation. We represent the space occupied by weeds as candidate line segments for spray and then construct a directed acyclic graph (DAG) that embraces the feasible nozzle paths among weeds. Based on the new scene representation, we formulate an optimal $K$ -nozzle assignment/motion planning problem and develop a binary linear programming-based algorithm to assign nozzles to the candidate line segments for optimal coverage. We built the system and conducted both simulation and field experiments. Evaluation on rough soil surfaces with artificial targets has shown that the lateral errors of herbicide spray are at sub-centimeter levels. Simulation results demonstrate that the proposed assignment algorithm can provide good coverage within the intra-row regions.