Publikation
Artificial Intelligence XL
Christoph Manss; Isabel Guenther; Florian Rahe (Hrsg.)
SGAI International Conference on Artificial Intelligence (AI-2023), Cambridge, United Kingdom, Vol. 14381, ISBN 978-3-031-47993-9, Springer Nature Switzerland, Cham, 2023.
Zusammenfassung
Intelligent weed management is crucial for sustainable and economical crop production. It optimizes agricultural machinery usage and reduces herbicide dependency. Advances in artificial intelligence and object detection algorithms have allowed for the use of cameras on land machines for real-time crop and weed detection. Robotic systems incorporate intelligent camera systems for weed identification and removal. However, training deep learning algorithms requires labor-intensive data cleaning and labeling. Weak supervision techniques offer a solution by reducing labeling efforts. This study explores inaccurate labeling for weed-crop differentiation in selective hoeing. A sensor-carrier with four cameras captures diverse images which are then labeled with inaccurate supervision. Multiple neural networks, including DeepLabV3, FCN, LRASPP, and U-Net, are trained using inaccurately labeled semantic masks. Results demonstrate reasonable performance while having less labeling time for developing smart weeding applications. For the considered data, the U-Net model exhibits the best mean intersection-over-union (IOU). This research emphasizes the potential of weak supervision and AI-based weed management for sustainable agriculture.