Abstract: Optical spectral reflectance and image analysis techniques were investigated as possible solutions to discriminate crop and weed plants. The range of pants included two brassica crop species, a cereal crop and eight weed species. Spectral signatures were obtained form optical reflectance measurement taken with a spectrophotometer in reflectance mode in the region between 700 and 1350 nm. Algorithms were developed based on multivariate statistical analysis of the plant reflectance spectra. By minimizing wavebands of interest for certain crop/weed combinations, better than 95 percent discrimination accuracy was obtained for only two or three waveband measures. Using filters at these wavebands it was possible to easily segregate corp from weed plants in images. Discrimination on the basis of leaf texture was investigated using textural signatures for whole leaves derived from a gray level co-occurrence matrix of nearest- neighbor pixel intensity. Textural features of leaves were expressed in the form of feature vectors comprising nine textural parameters extracted from the co-occurrence matrix. A numerical Bayesian classifier was used to classify leaves based on minimum distance between a mean feature vector determined form a training set and the test feature vector. A mean discrimination accuracy of 90 percent was achieved between al plant species and almost 100 percent separation was achieved between the crop and weeds. The results show that a combination of spectral imaging and texture analysis may provide a robust method of discrimination with potential for real time application. !9
展开▼
