Intelligent drones with vision technology for sustainable weed mapping in precision agriculture

This contribution outlines ongoing research in drone vision conducted at the Computational Intelligence Laboratory of the Department of Computer Science, University of Bari Aldo Moro, for drone-based weed mapping in precision agriculture. The research integrates multispectral imaging, deep learning, and model compression to develop efficient and deployable solutions. A key focus is on transfer learning strategies that enable pre-trained deep learning models on RGB datasets to operate effectively on multispectral imagery, addressing the scarcity of annotated agricultural data. Further work explores knowledge distillation techniques to compress complex segmentation networks into lightweight models suitable for real-time inference on drones and low-power edge devices, preserving high accuracy with minimal computational cost. A fully unsupervised framework—RoWeeder—is also introduced, which exploits crop-row detection to generate pseudo-ground truth for training segmentation models without manual annotations. Across these lines of research, experiments on the WeedMap dataset demonstrate robust weed detection under varying field conditions, advancing the applicability of drone-based systems in sustainable agriculture.