Multitemporal Hyperspectral Image Super-Resolution through 3D Generative Adversarial Network

The super-resolution of multitemporal hyperspectral imagery is considered, wherein a 3D generative adversarial network (GAN) is promoted and employed. Firstly, we put the SR process in a generative adversarial network (GAN) framework, so that the resulted high resolution HSI can keep more texture details. Secondly, the input of our method is of full bands due to 3D kernel exploited. Furthermore, a series of spatial-spectral constraints or loss functions are imposed to guide the training of our generative network so as to further alleviate spectral distortion and texture blur. The experiments on the houston datasets demonstrate that the proposed GAN-based SR method with the best generalization ability can yield very high quality results.