2.4.2. Support Vector Algorithms

The SVM andSVR are established machine learning methods that have been proven to deal well in situations with many features but a very limited number of samples [35]. This is a common situation in hyperspectral data analysis, and following it is a suitable approach for hyperspectral remote sensing as well as close range imaging. A critical point for the application of SVM and SVR is the selection of the hyper parameters Cost *C*, kernel parameter γ (SVM) or *C*, and complexity control ν (ν-SVR). They were selected by grid search combined with a cross validation. Grid points were 10−<sup>5</sup> ... 1010 for *C*, 10−<sup>8</sup> ... 10<sup>2</sup> for ν and 0.05 ... 0.50 for n. The optimization algorithm was the sequential minimal optimization SMO, and LIBSVM 3.18 with Matlab was used for as implementation [36].
