Performance of vegetation classification methods using synthetic multispectral satellite data

By including all available in-situ information from an area simultaneously imaged by the Landsat and Spot satellites, a vegetation map was constructed, and used to estimate the intensity distributions of five different vegetation classes in as many as ten spectral channels. The estimated class distributions were then used to generate synthetic multispectral data with the statistical properties of the real data. These synthetic data were used as input to some multispectral classifiers, including the classical Bayes classifier, a neural network classifier without contextual information, a neural network classifier with contextual information included as a functional link node, and a neural network classifier with contextual information as the intensity values of the neighbouring pixels. The results show that the neural network classifiers have a much less percentage of misclassifications than the Bayes classifier. By including contextual information of the pixels to be classified, the performance was found be significantly improved.