Algorithm 1 Algorithm of the proposed deep unsupervised fusion learning method.

Input: The observed LR-HS image $\mathbf{X}$ and HR-RGB image $\mathbf{Y}$

Output: Latent HR-HS image $\mathbf{Z}$

1: Sample ${\mathbf{z}}_{in}^0$ from uniform distribution with seed 0

2: for $i=0$ to max. iter. $(I)$ do

3:     Sample ${\mathbf{n}}_{(0,1)}^i$ from uniform distribution

4:     Perturb ${\mathbf{z}}_{in}^0$ with ${\mathbf{n}}_{(0,1)}^i$: ${\mathbf{z}}_{in}^i={\mathbf{z}}_{in}^0+\beta {\mathbf{n}}_{(0,1)}^i$

5:     $\widehat{\mathbf{Z}}=G_{\theta }({\mathbf{z}}_{in}^i,{\theta }^{i-1})$

6:     $\widehat{\mathbf{X}}=f_{DB}(\widehat{\mathbf{Z}})$

7:     $\widehat{Y}=f_C(\widehat{\mathbf{Z}})$

8:     Loss function: $\alpha {\beta }_1\left|\right|\mathbf{X}-\widehat{\mathbf{X}}{\left|\right|}_F^2+(1-\alpha ){\beta }_2\left|\right|\mathbf{Y}-\widehat{\mathbf{Y}}{\left|\right|}_F^2$

9:     Compute the gradients regarding $G_{\theta }$

10:     Update $\theta$ using the ADAM algorithm [47] as ${\theta }^i$

11: end for

12: $\mathbf{Z}=G_{\theta }({\mathbf{z}}_{in}^0)$