Evaluation of Hyperspectral Image Classification Using Random Forest and Fukunaga-Koontz Transform

Since hyperspectral imagery (HSI) (or remotely sensed data) provides more information (or additional bands) than traditional gray level and color images, it can be used to improve the performance of image classification applications. A hyperspectral image presents spectral features (also called spectral signature) of regions in the image as well as spatial features. Feature reduction, selection, and transformation has been a challenging problem for hyperspectral image classification due to the high number of dimensions. In this paper, we firstly use Random Forest (RF) algorithm to select significant features and then apply Kernel Fukunaga Koontz Transform (K-FKT), a non-linear statistical technique, for the classification. We provide our experimental results on AVIRIS hyperspectral image dataset that contains various types of field crops. In our experimental results, we have obtained overall classification accuracy around 84 percent for the classification of 16 types of field crops.