Hyperspectral and Multispectral Image Fusion Based on Unmixing-Like

The unmixing-based fusion methods mainly has the following two shortcomings. First, the high-resolution hyperspectral image (HR-HSI) is reconstructed using low-resolution hyperspectral spectral features (LR-HSI), and LR-HSI is the spatial degradation of the HR-HSI, which will lead to spectral distortion. Second, in the fusion method based on non-negative matrix decomposition, the solution of the objective function is sensitive to the initial value due to its non-convexity, which leads to the instability of the endmembers and the abundances. For these two problems, a hyperspectral image and multispectral image fusion method based on unmixing-like is proposed in this paper. Firstly, fuzzy c-means (FCM) is used to cluster the images, and the pixels closest to the cluster center are used as endmembers, which avoids the problem of solution instability caused by unmixing. Secondly, general regression neural network (GRNN) is used to learning the nonlinear relationship between HR-HSI and LR-HSI of the same scene in the spectral domain, which compensates for the endmember spectral distortion caused by spatial degradation. Finally, since the training data cannot cover all the spectrum types of the test data, the model output is optimized according to the test data. In this paper, the idea of unmixing-based fusion method is used and the endmembers of the LR-HSI is used to reconstruct the endmembers of HR-HSI. The reconstructed endmembers is combined with the sparse coefficient to obtain a high-spatial resolution image. The experiments used three sets of experiment data and the proposed fusion method is suitable for many datasets and achieved good robustness. The experiments show that the experimental results of Salinas data are 5.5%, 5.5% and 1.6% lower in SAM, RMSE and ERGAS indexes compared with the sub-optimal method. In the Cuprite data, the values decrease by 1.3%, 3.9% and 3.8%. In Indian Pines, they are down 1.7%, 4.0% and 3.9% respectively. With respect to Pavia Center data, the indexes of SAM and ERGAS are reduced by 2.9% and 8.5% respectively in the case of bicubic interpolation. When bilinear interpolation is used, the values are reduced by 3.5% and 3.4% respectively. Therefore, the algorithm in this paper can effectively improve the spatial resolution and maintain the spectral information at the same time. © 2020, Beijing China Science Journal Publishing Co. Ltd. All right reserved.