Sensor for weed detection based on spectral measurements

A sensor is proposed here which distinguishes between crop and weed based on their different spectral reflectances. The sensor is built upon an imaging spectrograph. We chose this spectral reflectance sensor because of the fast spectral imaging and possible high spatial and spectral resolution. Parameters like the angle-of-view and the quality of the optics were optimized for maximal performance within reasonable cost. Classification success rates depend not only on spatial and spectral filtering, both characteristics of the device, but also on the number of wavelengths and the crop itself. Under controlled conditions, corn and sugar beet can be separated from weed with a success rate of at most 90, respectively 80%. Herbicide savings which depend on weed density, the nozzle activation frequency and the spray resolution (width), are maximal with the MLNN classifier.