6533b7dafe1ef96bd126e83f

RESEARCH PRODUCT

Modélisation d'images agronomiques - application a la reconnaissance d'adventices par imagerie pour une pulvérisation localisée

subject

description

Plant (crop and weed) identification is a very active field of research in agriculture since the reinforcement of European laws about pesticide applications for a site-specific management of spraying practices. A new crop/weed simulation model was developed to allow the evaluation of crop/weed spatial identification methods from imaging. Considering multiples agronomic parameters – crop location, weed infestation rate, weed spatial distribution – the first step of this model allows the simulation of an infested crop field. Then, in a second step, a world to camera transformation is applied to allow every kind of picture (with or without perspective effect). The validation of this model was performed using statistical tests comparing a real image to its homologous virtual one. New crop/weed discrimination algorithms based on the Hough Transform to detect crop rows were also developed. Three methods, using the crop row information and based on a blob-coloring, an edge estimation or a probabilistic classification were exhaustively tested using this model. Results show very good performance of these methods with correct average classification rate of 90% and up to 98% under special conditions. A spectral approach was also explored for the model in order to overcome the limitations imposed by spatial algorithms. Crop and weed plant patterns are now in 3D to allow the calculation of the bidirectional reflectance (BRDF) of plants and soil based on PROSPECT and SOILSPECT models. We also discussed the transformation of a reflectance spectrum into a RGB color, the simulation of optical filter effects and the creation of multispectral images.