Multi-class support vector machine classifier applied to hyper-spectral data

The availability of sufficient training data is a continuing challenge for most real world classification problems. This is especially true for hyperspectral processing since there are a very limited number of high quality hyperspectral sensors. In general, discrimination techniques require a large amount of training data in order to produce reliable probability distributions. Unfortunately, there is usually not a statistically significant amount of real data to adequately describe the various object classes. Given this limitation, we propose a novel algorithm approach based on support vector machines (SVM). Support vector machines are a type of learning machine based on statistical learning theory developed by Vapnik (1982, 1995). SVM overcome the limitations of traditional discrimination approaches that seek to minimize risk based solely on training data. SVM accomplishes this through its ability to "generalize errors" which has been shown to be more robust than traditional risk approaches. This capability to generalize makes SVM ideal for real world problems. This paper will present a support vector machine algorithm applied to a multiclass hyperspectral scenario using data from the AVIRIS sensor.