Plant Classification Combining Colour and Spectral Cameras for Weed Control Purposes

Weed plant detection and classification is a difficult task for any computer vision system. Previous studies show promising results with either colour camera or spectral imaging solutions. However, typical colour camera solutions have found it hard to deal with overlapping leaves, and spectral solutions often lack in the spatial resolution required for accurate leaf level detection. In this paper a novel system for weed detection and classification is presented using both low-cost RGB (Red, Green, Blue) colour and spectral (400 - 1000 nm) cameras combining the strengths of these individual technologies. The system presented performs accurate leaf level classification and is capable of identification at 97.6% with non-overlapping full leaves in laboratory under controlled lighting conditions. Plant leaf samples from 6 different plant types were used. With dedicated hardware and optimized software the system should be capable of at least 5 km/h real-time operation in field conditions.