Global Urban-Rural Floods Dataset

This deposit provides the data, code, and figures supporting a global analysis of urban versus matched-rural flood occurrence. It includes the Google Earth Engine workflow used to derive flood occurrence from satellite observations, frozen analysis tables, a one-command Python pipeline that regenerates all 37 code-generated analytical panels, and the resulting figures and numerical source tables. Method Flood occurrence is measured as the fraction of valid satellite observations in which a pixel is classified as water. The workflow combines 10 m Sentinel-2 Dynamic World observations with supplementary geospatial layers and the following physical and contextual filters: Dynamic World: water is identified using the label and probability layers (water > 0.5); flooded vegetation (label == 3) is excluded from the valid-observation denominator. ESA WorldCover: permanent water, wetlands, and mangroves are masked (classes 80, 90, and 95). FABDEM: terrain with slope greater than 5° is removed. GHSL: built structures are masked using P2023A built-characteristics values ≥ 11. The primary Earth Engine method uses Sentinel-2 L1C for 2018 and L2A for 2019–2025. Mean flood occurrence is calculated over the deposited 2022 urban and matched-rural boundary system. The included workflows support annual and complete-period analyses, as well as cloud-threshold, persistent-water-cap, boundary-year, and Sentinel-2 collection sensitivities. Primary Data Product The canonical analysis table contains 81,319 cities, with: flood tendency; aggregate and annual urban/rural flood metrics for 2018–2025; land-cover characteristics; elevation; coastal/inland classification; Global North/South classification; urban area; and population. For provenance, the canonical table has the following column-level lineage: Unsuffixed aggregate metrics come from the retained all-years Sentinel-2 L1C source. Annual 2018 fields use the Sentinel-2 L1C cloud-35 source. Annual 2019–2025 fields come from the Sentinel-2 L2A master table. All three source tables are included so the canonical table can be rebuilt and audited column by column. Reproducing the Analysis Create the conda environment from environment.yml: conda env create -f environment.yml Run the complete reproduction pipeline: conda run -n gufd-final-release python scripts/reproduce_all.py This validates the reported numerical results and regenerates all 37 code-generated analytical panels. Individual panels and Google Earth Engine workflows can also be run separately. See the following files for complete instructions and provenance: README.md docs/PANEL_PROVENANCE.csv code/gee/README.md Reproduction Limits Three additional manuscript visuals are included as reference-only artwork because their original editable or raw visual inputs were not recoverable: Figure 1a — workflow layout; Extended Data Figure 9 — city-example rasters and paired polygons; and Figure S6 — raw satellite-scene layers. Their final submitted artwork is included under figures/manuscript_composites/. See docs/LIMITATIONS.md for details.