Classification of multi-spectral image data considering non-Gaussian distribution and inter-pixel class dependency

Classification can partition the feature space of a multi-spectral image into separated regions corresponding to specified classes. Pixels close to one or some of the training examples can be classified easily. Else pixels may be either different from the examples or near the class boundary, and classification by simply evaluating spectral characteristics can be far from accurate or even impossible. By considering inter-pixel class dependency as auxiliary information in spectral classification, this paper proposes a new classifier to improve the accuracy in classifying multi-spectral image data. This new classifier evaluates the spectral characteristics of each pixel first to categorize it as a classifiable or a spectrally uncertain pixel. Inter-pixel class dependency is then added to classify the uncertain pixels. The class similarity between spatially adjacent pixels of remotely observed multi-spectral image motivates the method of using inter-pixel class dependency. A spatial filter with pre-specified parameters is used to describe the dependency. The distributions of the probability density functions for classification are not assumed to be normal and a non-parametric method is adopted to estimate the class-conditional probabilities. Comparison with those using Gaussian maximum likelihood and back-propagation neural networks is demonstrated and excellent simulation results are obtained.