Multispectral classification of grass weeds and wheat (Triticum durum) using linear and nonparametric functional discriminant analysis and neural networks

Field studies were conducted to determine the potential of multispectral classification of late-season grass weeds in wheat. Several classification techniques have been used to discriminate differences in reflectance between wheat and Avena sterilis, Phalaris brachystachys, Lolium rigidum and Polypogon monspeliensis in the 400-900 nm spectrum, and to evaluate the accuracy of performance for a spectral signature classification into the plant species or group to which it belongs. Fisher's linear discriminant analysis, nonparametric functional discriminant analysis and several neural networks have been applied, either with a preliminary principal component analysis (PCA) or not and in different scenarios. Fisher's linear discriminant analysis, feedforward neural networks and one-layer neural network, all showed classification percentages between 90% and 100% with PCA. Generally, a preliminary computation of the most relevant principal components considerably improves the correct classification percentage. These results are promising because A. sterilis and L. rigidum, two of the most problematic, clearly patchy and expensive-to-control weeds in wheat, could be successfully discriminated from wheat in the 400-900 nm range. Our results suggest that mapping grass weed patches in wheat could be feasible with analysis of real-time and high-resolution satellite imagery acquired in mid-May under these conditions.