3D Convolutional Neural Networks with Image Fusion for Hyperspectral Image Classification

Image fusion can extract redundant information of multiple images into one image, and the goal of image fusion is to better apply to classification tasks. Convolutional neural networks have proved to be an effective way for accuracy of classification. However, fusion and classification usually considered separately. In this paper, we design a ‘fusion-classification networks’, and introduce image fusion technology and 3D convolutional neural networks (3D CNNs) into HSI classification. In the proposed method, the fusion process is guided by the classification result, and the classification accuracy is improved by the fusion process. Image fusion technology is performed on spectral bands to exploit the redundancy information of HSI, and 3D CNNs are applied on the fused image to extract more robust spectral-spatial features. The proposed method is tested on two datasets. Its outstanding performance is validated in comparison with other state-of-the-art approaches.