Detection of weeds using image processing and clustering

Knowledge about the distribution of weeds in the field is a prerequisite for site-specific treatment. Optical sensors make it possible to detect varying weed densities and species, which can be mapped using GPS data. The weeds are extracted from images using image processing and described by shape features. A classification based on the features reveals the type and number of weeds per image. For the classification only a maximum of 16 features out of the 81 computed ones are used. Features are used, which enable an optimal distinction of the weed classes. The selection can be done using data mining algorithms, which rate the discriminance of the features of prototypes. If no prototypes are available, clustering algorithms can be used to automatically generate clusters. In a next step weed classes can be assigned to the clusters. Such a procedure aids to select prototypes, which is done manually. Classes can be identified, that are distinct in the feature space or which are overlapping and therefore not well separable. Clustering can be used in some, less complex cases to establish an automatic procedure for the classification. Weed maps are generated using the system. These are compared to the results of a manual weed sampling.