Support vector classification applied to wheat and ryegrass spectral signatures discrimination: dimensionality reduction analysis

Abstract

Spectral signatures offer valuable insights for distinguishing between crops and weeds because they provide the unique light reflection patterns specific to each plant species. Remote sensing classification algorithms have shown strong potential for weed detection during crop preemergence, benefiting from the clear spectral differences between weeds and bare soil. However, achieving effective site-specific and real-time weed management after crop emergence remains challenging. This study explores this issue through controlled pot trials, evaluating the spectral differentiation between wheat and ryegrass at various phenological stages using a support vector machine classifier. Spectral data were analyzed at multiple resolutions—1 nm, 5 nm, 10 nm, and the discrete five-band configuration of the MicaSense RedEdge-MX multispectral camera. Classification performance was evaluated using confusion matrices and key metrics such as accuracy, precision, recall, and F1 score. The results consistently indicated a strong discrimination capability across configurations, with the MicaSense bands achieving optimal performance. These findings suggest that strategic band selection balances noise reduction while preserving critical spectral features. Scaling up these insights to real production situations could pave the way to developing site-specific real-time applications for weed management with emerged crops.