A multi-sensor co-registration benchmark: Sentinel-2, Landsat-8/9 and NAIP image pairs over 60 global sites (SCDF legs B1-B5)

Benchmark imagery of the paper Self-Calibrating Dense Displacement Fields for Reliable Co-Registration of Large Optical Satellite Imagery (S. Sun, Z. Wang, S. Salati, J. Zhang, H. Wang, X. Ma; arXiv:2608.22300): raw image pairs of the paper's five benchmark legs over 60 sites worldwide, and the script that regenerates the paper's synthetic constructions (A1–A4). The constructed pairs themselves are not included. Legs (reference / moving, sites): B1 Sentinel-2 2024-07 / Sentinel-2 2024-01, 25 B2 Sentinel-2 2024-07 / Sentinel-2 2023-07, 25 B3 Sentinel-2 2024-07 / Sentinel-2 2019-07, 25 B4 Sentinel-2 10 m / Landsat-8/9 30 m, 25 B5 Sentinel-2 10 m / NAIP aerial 0.6 m, 21 Files: one ZIP archive per leg, plus manifest.csv, checksums.sha256, stac_provenance_b4_b5.csv, sites.geojson and make_constructions.py. See README.md for the contents and usage. Software: the registration method is released as the geocoreg Python library at https://github.com/TitorX/geocoreg. Attribution: contains modified Copernicus Sentinel data (2019–2024). Landsat imagery courtesy of the U.S. Geological Survey. NAIP imagery provided by USDA Farm Service Agency. Imagery was accessed through the Microsoft Planetary Computer.