Abstract

The increasing availability of high-resolution satellite imagery has driven advances in 3D reconstruction techniques for the generation of Digital Elevation Models (DEM). Recent research focuses on opportunistic stereo, using sophisticated techniques like Neural Radiance Fields and Gaussian Splats, that are able to exploit a collection of multi-date images of the same site. These techniques give optimal results, but they are computationally expensive, so in practice, they are only used for small regions of interest. In contrast, quasi-simultaneous stereo products are routinely acquired for large-scale mapping, needing a focus on efficiency, robustness and scalability in their processing. This paper introduces s2p-hd, a binocular stereo pipeline designed for high-throughput processing of same-date satellite imagery. Building upon the open-source~s2p pipeline, s2p-hd adds several key improvements that enhance its performance and robustness, tuned for same-date stereo imagery. These include a refined disparity range estimation leveraging reference models and multiscale analysis, the adaptation of a highly optimized GPU-based Semi-Global Matching (SGM) algorithm, and enhanced rectification and tiling strategies. We benchmark s2p-hd against standard stereo pipelines and show that it outperforms them both in accuracy and processing speed, making it a powerful tool for generating high-quality DEMs from large-scale optical satellite imagery, while balancing precision and computational efficiency.