In this paper, we successfully apply masked image modeling (MIM) to the dehazing process for satellite images, introducing a novel dehazing method. Initially, we investigate why MIM does not effectively function as a self-supervised learning method for low-level vision tasks and fails to yield improved performance. Subsequently, we propose two solutions to address this issue. Furthermore, we introduce an augmentation technique that enhances both locality and non-locality in puzzle images through jigsaw transformations, resulting in improved accuracy. Experimental results show that our method outperforms other state-of-the-art methods, including approaches based on visual transformers.