Infra-Bench Classification (Infra-Bench CLS) Imagery Dataset

Infra-Bench Classification (Infra-Bench CLS) pairs facility-scale critical-infrastructure assets identified in OpenStreetMap with co-located Sentinel-1 and Sentinel-2 imagery, so that Earth-observation foundation models can be compared on a common classification task across seven continental regions and four infrastructure sectors. Each asset is represented by one 600m x 600m tile centred on its location, stacking seven Sentinel-2 L2A optical bands with the two Sentinel-1 RTC polarizations at 10m resolution. Labels are weak, in the sense that classes derive from OpenStreetMap tags rather than from manual inspection. The archive contains two parts. The benchmark comprises 28 "cells" named dataset___v1_1k, covering seven regions crossed with four sectors, each targeting 1,000 tiles drawn by proportional class-stratified sampling, for 18,756 tiles in total across a 13-class taxonomy. Additionally, the substation datasets, dataset__stac_v1, holds every deduplicated substation per region within the energy sector and without a per-cell cap. Each cell (e.g., Africa x energy) is a zip archive holding an images/ directory of .npy tiles, a manifest.json of per-tile metadata, and a summary.csv. Asset locations come from a Geofabrik OpenStreetMap snapshot dated 2026-05-26, and imagery from the Microsoft Planetary Computer, selecting the least-cloudy scene available for each asset within a 2021 to 2024 window. Note that the image_date field in these manifests records the end of that search window rather than the acquisition date of the scene used, so it reads 2024-12-31 for every tile. Curation and evaluation code, including the fixed train, validation and test split used in the paper, is available at github.com under the MIT License. Released under the Open Database License (ODbL) 1.0. Attribution: Contains information from OpenStreetMap Planet PBF (Geofabrik snapshot 2026-05-26). (c) OpenStreetMap contributors, licensed under ODbL 1.0. Imagery derived from modified Copernicus Sentinel-1 and Sentinel-2 data.