Fast feature selection methods for classification of hyperspectral images

With development of hyperspectral imaging, it is possible to identify and classify land cover with more details in remote sensing applications. Selection of a minimal and efficient subset from the huge amount of features is an important challenge for classification problems. Almost all approaches for feature selection, which represented in literature, involve a search algorithm for selection of the best candidate from possible solutions and are very time consuming. We propose two feature selection methods in this paper that need no search algorithm. The methods select the efficient subset of features by using a simple calculation of standard deviation and mean values. Thus, the proposed methods are run fast. The experimental results using three different hyperspectral images demonstrate the high speed and reasonable performance of proposed methods in comparison with sequential forward selection (SFS). We select the SFS algorithm because it is a simple and suboptimal technique.