Home / Works / Cebu City Green Space and Flood-Prone Screening

Open the interactive map   Methodology and limits

Overview

A first-pass screening of Cebu City that asks where people live with little green space, on flat low ground near drainage, and far from schools and health facilities. It is built only from open global data and a pipeline that re-runs from one config file: a Sentinel-2 scene, the Copernicus 30 m elevation model, OpenStreetMap, Google Open Buildings and NAMRIA river lines. It is a screening tool, not a hazard map.

How it works

Five open datasets become land cover, a flood proxy and service distances, then one 100 m grid.

InputStepDecision / optionData storeOutput

What it found

• Land cover inside the 309 km² city boundary: 77.8% green space, 12.5% non-vegetated (built-up or bare), 4.2% water, with 5.4% masked as cloud or shadow.
• A flood-proneness proxy (slope of 2° or less and no more than 5 m above nearby drainage) flags 12.8 km², 4.1% of the boundary, mostly on the coastal plain and river valleys.
• OpenStreetMap lists about 40,000 buildings in the city, while Google Open Buildings finds 183,000 at a confidence of 0.70 or more. Only about a quarter of those footprints touch an OSM building, so OSM is far from complete here.
• The three criteria point at different places, so a single weighted score is a policy choice. Among populated 100 m cells, 1,482 cells (57,466 building footprints) are both green-poor and flood-prone, and only 3 cells meet all three criteria.

How I checked it

• Land cover: thresholds chosen on alternating 2 km blocks and scored on the other blocks against ESA WorldCover: mean F1 0.82 on the held-out blocks, up from 0.68. WorldCover is itself a model, so this is agreement, not accuracy.
• Sensitivity: across 19 weight combinations the top-10% cells overlap the equal-weight ranking by a median of 0.59 (0.08 at worst), and 62% of the top-10% cells are empty ground, so the map scores only populated cells.
• Data checks: OSM roads agree with an independent count to 0.3%; DEM-derived streams follow the major NAMRIA rivers (80% within 150 m).
• Code: numbered pipeline scripts and 28 automated tests.

Not done yet

• The flood proxy has not been compared with real flood records, and its thresholds are round numbers.
• The land-cover map has not had a by-eye accuracy check, and the school and health layers have not been cross-checked against a second source.
• The code lives in a private repository for now.

Credits