Abstract: Large-scale remote sensing images, including both satellite and aerial photographs, are widely used to render terrain scenes in real-time geographic visualization systems. Such systems often require large memories in order to store fine terrain details and fast network speeds to transfer image data, if they are built as web applications. In this paper, we propose a progressive texture compression framework to reduce the memory and bandwidth cost by compressing repeated content within and among large-scale remote sensing images. Different from existing image factorization methods, our algorithm incrementally find similar regions in new images so that large-scale images can be more efficiently compressed over time. We further propose a descriptor, the Gray Split Rotate (GSR) descriptor, to accelerate the similarity search. The reconstruction quality is finally improved by compressing residual error maps using customized S3TC-like compression. Our experiment shows that even with the error maps, our system still has higher compression rate and higher compression quality than using S3TC alone, which is a typical compression solution in most existing visualization systems.