The Robustness of Natural Image Priors in Remote Sensing: A Zero-Shot VAE Study

Zhenyuan Chen, Feng Zhang

Abstract: This paper explores the robustness of variational autoencoders (VAEs) pre-trained on natural image data, such as ImageNet, when applied to the remote sensing domain in a zero-shot manner. We investigate whether these natural image priors embedded in standard VAEs can serve as effective compressors and reconstructors for satellite images, even when applied in a different manner across various settings compared to natural cases. Our study evaluates several state-of-the-art VAE architectures across multiple remote sensing categories and reconstruction metrics to demonstrate their potential. See code at \url{https://github.com/Bili-Sakura/VAEs4RS}.

Submission Number: 3