Hyperspectral Image Classification Based on Multiscale Spatial Information Fusion

In hyperspectral image (HSI) classification, the combination of spectral information and spatial information can be applied to enhance the classification performance. In order to better characterize the variability of spatial features at different scales, we propose a new framework called multiscale spatial information fusion (MSIF). The MSIF consists of three parts: multiscale spatial information extraction, local 1-D embedding (L1-DE), and information fusion. First, spatial filter with different scales is used to extract multiscale spatial information. Then, L1-DE is utilized to map the spectral information and spatial information at different scales into 1-D space, respectively. Finally, the obtained 1-D coordinates are used to label the unlabeled spatial neighbors of the labeled samples. The proposed MSIF captures intrinsic spatial information contained in homogeneous regions of different sizes by multiscale strategy. Since the spatial information at different scales is processed separately in MSIF, the variance of spatial information at different scales can be reflected. The use of L1-DE reduces computational cost by mapping high-dimensional samples into 1-D space. In MSIF, the L1-DE and information fusion are used iteratively, and the iterative process terminates in a finite number of steps. The algorithm analysis demonstrates the effectiveness of the proposed method. The experimental results on four widely used HSI data sets show that the proposed method achieved higher classification accuracies compared with other state-of-the-art spectral-spatial classification methods.