Applying boosting for hyperspectral classification of ore-bearing rocks

Hyperspectral sensors provide a powerful tool for non-destructive analysis of rocks. While classification of spectrally distinct materials can be performed by traditional methods, identification of different rock types or grades composed of similar materials remains a challenge because spectra are in many cases similar. In this paper, we investigate the application of boosting algorithms to classify hyperspectral data of ore rock samples into multiple discrete categories. Two variants of boosting, GentleBoost and LogitBoost, were implemented and compared with Support Vector Machines as benchmark. Two pre-processing transformations that may improve classification accuracy were investigated: derivative analysis and smoothing, both calculated by the Savitzky-Golay method. To assess the performance of the algorithms over noisy data, white Gaussian noise was added at various levels to the data set. We present experimental results using hyperspectral data collected from rock samples from an iron ore mine.