Rethinking Pan-Sharpening in Closed-Loop Regularization.