Bi-Temporal Remote Sensing Image Fusion Via Semi-Coupled Low-Rank Tensor Approximation

Fusion of hyperspectral (HS) and multispectral (MS) images at single time has been well studied, but the problem of fusing these images acquired from different dates still remains to be solved. Up till now, current methods fail to establish an efficient temporal mapping extraction scheme while make use of the advantage of tensor-based model. To deal with this issue, a novel bi-temporal HS-MS fusion method called Semi-coupled Low-rank Tensor Approximation (S-LRTA) is proposed. The method firstly employs Tucker decomposition to make the fusion task a factor estimation problem. Then it captures the natural low-rank property of hyperspectral image (HSI) with a sparse constraint on the core tensor. Particularly, a Hadamard-product based temporal variability descriptor is blended into the Tucker model to extract the temporal relationship which is the crucial difficulty in bi-temporal fusion problem. Lastly, an efficient Block Coordinate Descent (BCD) based optimization scheme is developed to solve the objective function. Experimental results demonstrate the superiority of the proposed method compared with state-of-the-art methods.