Reducing hyperspectral data dimensionality using random forest based wrappers

The random forest algorithm has been widely used for classification of hyperspectral data. To improve model interpretation and classification, the random forest algorithm is often combined with feature selection algorithms. It is within this context that we explore the utility of three random forest wrappers to compute an optimal subset of wavebands to discriminate healthy and stressed Pinus radiata seedlings. The Boruta algorithm provided the best classification results using a subset of 17 wavebands of an original 1 769 wavebands. This study demonstrated the value of using wrappers embedded within the random forest algorithm for classification of high dimensional data. in particular, this study highlights the application of the Boruta algorithm for discriminating healthy and stressed P. radiata seedlings.