Résumé
In maize fields, weeds often appear in patches. Herbicide use can be reduced by up to 75% if their application is limited to weed infested areas (spot spraying). This can be performed if infested areas can be located. This paper reports on work to assess the potential of UV-induced fluorescence in discriminating between maize and weeds. The experimental program covered three years where plants were grown and measured under contrasting conditions. Plants were grown either in a growth chamber under artificial lighting or in a greenhouse under natural sunlight. Measurements were performed either in the greenhouse under natural sunlight or in a growth chamber under standardized lighting conditions. Several factors were considered: weed type (monocots or dicots), leaf side and position of measurement on the leaf. Plant fluorescence was induced at wavelengths around 328 nm and fluorescence measured between 400 and 760 nm. The discrimination between monocots and dicots based on the blue-green fluorescence yielded robust models with a classification error between 1.3 and 4.6%. From the partial least-square discriminant model, two large bands were chosen in the blue-green fluorescence: 400-425 nm and 425-490 nm. A linear discriminant analysis based on these two bands provided robust inter-year results with classification errors from 1.5% to 5.2%. Discriminating between maize (a monocot) and monocot weeds proved more difficult. Robust discrimination can be achieved based on properly normalized chlorophyll fluorescence signal and may require recalibration using a small sample of maize plants each year or each time a new field is being sprayed. For maize-monocot discrimination, the error rate for robust discrimination was 15%. Globally, our results indicate that weed detection based on UV-induced fluorescence could lower the use of herbicides in maize fields and yield economical and environmental benefits.