Active learning system for weed species recognition based on hyperspectral sensing

Weeds have a devastating impact in crop production and yield in general. Current practice uses uniform application of herbicides leading to high costs and degradation of the environment and the field productivity. Site-specific treatments can be regarded as solutions either for reducing inputs or enable alternative non-chemical treatments. However, site-specific treatment needs accurate targeting through sensing. A new machine learning method is proposed, which discriminates between crop and weed species relying on their spectral reflectance differences. Spectral features were extracted from a hyperspectral imaging system that was mounted on a robotic platform. The proposed machine learning method suggests active learning by combining novelty detection and incremental class augmentation. Novelty detection was based on one-class classifiers constructed by neural networks. Best results for the active learning were obtained for the one-class MOG (mixture of Gaussians) and one-class SOM (self-organising map) classifiers when compared with one-class support vector machines and the auto-encoder network. The SOM and MOG performance in crop recognition was found to be 100% and 100% respectively. The recognition performance for different weed species varied between 31% and 98% (MOG) and 53%-94% (SOM).