Extreme learning machine (ELM) is of great interest to the machine learning society due to its extremely simple training step. Its performance sensitivity to the number of hidden neurons is studied under the context of hyperspectral remote sensing image classification. An empirical linear relationship between the number of training samples and the number of hidden neurons is proposed. Such a relationship can be easily estimated with two small training sets and extended to large training sets to greatly reduce computational cost. The kernel version of ELM (KELM) is also implemented with the radial basis function kernel, and such a linear relationship is still suitable. The experimental results demonstrated that when the number of hidden neurons is appropriate, the performance of ELM may be slightly lower than the linear SVM, but the performance of KELM can be comparable to the kernel version of SVM (KSVM). The computational cost of ELM and KELM is much lower than that of the linear SVM and KSVM, respectively.

This study explores the use of the graphics processing units (GPUs) for performing the two-dimensional discrete wavelet transform (DWT) of images. The study of fast wavelet transforms has been driven both by the enormous volumes of data produced by modern cameras and by the need for real-time processing of these data. With the emergence of general computing on GPUs, many time-consuming applications have started to reap the associated benefits. In the implementation of a GPU-based DWT, two approaches are used according to the published works, which are the row–column (RC) approach and the block-based (BB) approach. Most state-of-the-art techniques are based on the RC approach, which utilises the parallelism between different rows and columns; few works are based on the BB approach, which explores the parallelism between different blocks of the image. Although easy to implement, resource usage of the RC approach is usually related to the image size. Another shortcoming of the RC approach lies in the fact, according to the author's analysis, that more global memory access is required. The authors thus select the BB approach in this study. Experiment results show that the proposed BB approach outperforms the RC approach, being 99× faster than a native CPU implementation for 4096 × 4096 images.