Hyperspectral Image Classification using Multi-channel Lightweight CNN Ensemble Framework Based on Band Clustering

This paper presents an investigation into the challenging task of hyperspectral image (HSI) classification. The work is inspired by the observation of recent success of deep learning, especially regarding their application in improving the performance of HSI classification. However, applying deep convolutional neural networks to HSI classification requires many parameters to be trained, incurring significant computational costs. To address this problem, an effective multi-channel lightweight CNN ensemble framework is proposed here, where each channel takes a different band of the training samples by a novel band clustering and selection strategies. After training the individual networks, a novel weighted voting strategy is applied to make a combined ensemble decision over the resulting set of classifications. As such, this work organically integrates the merits of ensemble learning and deep learning. Experimental results on two hyperspectral airborne images demonstrate that the proposed method outperforms the state-of-the-art deep learning methods, with higher classification accuracy and relatively shorter training time.