Supervised Multispectral Classification: The Multilayer Perceptron versus the Maximum Likelihood Classifier

Project Leader: Dr. Klaus Steinnocher

End: 12.2001

In this thesis a linear (maximum likelihood) and a non-linear (neural net) classifier are compared for the analysis of satellite images. Besides examining a standard per-pixel procedure, the usability of these classifiers for providing thematic sub-pixel information is examined. This is done by classifying an artificial data set, derived from the mean-values for each class. In a next step it will be examined whether the results can also be confirmed by examining the original classified data. It is expected that the use of sub-pixel information allows a better interpretation of the land cover present in the study area.

Based on test data the results of the linear and non-linear per-pixel classifications are comparable. A visual comparison shows notable differences for some classes, especially in areas that were not part of the training data. Here the performance of the neural net is superior as fewer misclassifications seem to occur. In order to analyse these differences, a statistical test was performed on the training data. It could be shown that the distribution of the data has a significant influence on the accuracy of the classification. As both classifiers are based on different assumptions concerning the distribution of the data, the whole feature space was classified in order to visualise how classes are separated. In order to examine the limits of a sub-pixel classification, an artificial data set based on the mean values of the training data, simulating a linear mixture between all two-class combinations, was created. The classification of this data set makes it possible to visualise which class mixtures can be separated by which classifier, which class mixtures cannot be separated at all and which mixtures leads to so-called systematic misclassification. The results confirm the hypothesis that as long as the data is normally distributed, both classifier give similar results while the neural net is superior when the data is not normally distributed. The visualisation of the artificially mixed classes allows a better interpretation of the per-pixel classification as now apparently wrongly classified pixels can be related to the presence of mixed pixels.