As single-layer feed-forward neural networks, extreme learning machine (ELM) has recently been used with success for the classification of hyperspectral images (HSIs). However, the results of pure pixel-wise spectral classifiers often appear very noisy with limited training samples. To further impro