Nowadays, the use of hyperspectral sensors has been extended to a variety of applications such as the classification of remote-sensing images. Recently, a spectral–spatial classification scheme (ELM-EMP) based on Extreme Learning Machine (ELM) and Extended Morphological Profiles (EMPs) computed using Principal Component Analysis (PCA) and morphological operations has been introduced. In this work, an efficient implementation of this scheme over commodity Graphics Processing Units (GPUs) is shown. Additionally, several techniques and optimizations are introduced to improve the accuracy of the classification. In particular, a scheme using an ELM classifier based on kernels (KELM) and EMP is presented (KELM-EMP). Similar schemes adding a spatial regularization process (KELM-EMP-S and ELM-EMP-S) are also proposed. Moreover, two PCA algorithms have been compared in both accuracy and speed terms. Regarding the GPU projection, different techniques and optimizations have been applied such as the use of optimized Compute Unified Device Architecture (CUDA) libraries or a block-asynchronous execution technique. As a result, the accuracy obtained by the two proposed schemes (ELM-EMP-S and KELM-EMP-S) is better than for the original scheme ELM-EMP and the execution time has been significantly reduced.
Keywords: Remote sensing, classification, hyperspectral data, Extreme Learning Machine (ELM), Principal Component Analysis (PCA), Extended Morphological Profiles (EMP)