K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 3.4

Application of multi-spectral image processing in agriculture: a weed sensor

L. Van Gool, P. Wambacq, F. Feyaerts, P. Pollet

Trends in modern agriculture towards precision farming call for sophisticated data processing techniques and sensors. The goal of this research is to develop technology to diminish the use of chemicals, which is desirable both from an economic and environmental point of view. The key to success of plant specific spraying (spray only when needed) is the ability to discriminate between crop and weed.

In this work, algorithms were developed for online measuring and classifying samples based on the different spectral reflectance of crop and weeds. By diffraction of the incoming light, reflectance spectra of a line are projected on a monochrome CCD camera. Robust - online - reflectance measurements are gathered and fed to a classification algorithm. Field measurements were carried out combining multi-spectral reflectance and structural information (crop rows). Under those circumstances, sugar beet can be discriminated from weeds with an accuracy of 86%. Possible herbicide reductions, depending on spray resolution and weed density, can be as high as 80%. Weed hit rates are at least as high as 91%. Minimizing the overall cost (spraying cost, yield losses by crop stress, ...) by introducing cost functions could improve these figures even more. Extension of this technique to discriminate between different kinds of weeds will enable selective treatment on monocotyls and dicotyls. This technique also has potential for other applications that need to discriminate based on colour features.