How It Works
Tulay Pinoy uses free Sentinel-2 satellite imagery and spectral change detection to look for visible construction at completed Philippine DPWH project sites, starting with flood control, the category at the centre of the 2025 infrastructure-spending review and the one whose footprints 10m imagery can actually resolve. Each completed site reads as construction visible or not visible from space. It is an open-source tool: the same pipeline runs on any country's infrastructure data.
The Core Insight
When humans build things, the Earth's surface changes in predictable ways:
Spectral Indices
- Bands
- Sentinel-2 B11 and B8
- Range
- -1.0 to +1.0
- Detection threshold
- +0.10 increase to detect new construction
- High values
- Built-up surfaces (concrete, asphalt, roofing)
- Low values
- Vegetation, water
- Bands
- Sentinel-2 B8 and B4
- Range
- -1.0 to +1.0
- Detection threshold
- -0.15 decrease to detect vegetation clearing
- High values
- Dense vegetation cover
- Low values
- Bare soil, built-up areas
- Bands
- Sentinel-2 B11, B4, B8, B2
- Range
- -1.0 to +1.0
- Detection threshold
- +0.10 increase as construction corroboration
- High values
- Exposed soil, excavation sites
- Low values
- Vegetation, water
Classification Logic
| Classification | Condition | Interpretation |
|---|---|---|
| Construction Detected | NDBI > +0.10 AND NDVI < −0.15 | Strong evidence both conditions met |
| Vegetation Cleared | NDVI < −0.15 AND NDBI ≤ +0.10 | Land cleared but no built-up signal yet |
| Partial Construction | NDBI > +0.10 AND NDVI ≥ −0.15 | Built-up increase without full vegetation removal |
| No Clear Change | Below all thresholds | Inconclusive: weak or no signal at 10m resolution |
| No Construction Visible | Completed + built-up index flat or falling (top of absence rank) | No construction visible where a finished project should show it. A prompt to look, not a claim |
One honest caveat up front: as a plain yes/no detector this method is far too blunt. Run as a binary "built or not" test on completed flood-control projects, our calibration against real Sentinel-2 imagery reads roughly two-thirds to four-fifths of them as absent, because most flood-control structures (concrete on already-bare riverbanks) produce a weak spectral signal, not because the work is missing. So we do not use that raw call. Instead every assessed project gets a continuous absence score from how flat or falling its built-up index is, and only the strongest tail (completed projects where the built-up index actually held flat or dropped) is shown in red as no construction visible. That cut is deliberately conservative: at about 2 percent of assessed sites it sits well below the rate the government's own Independent Commission for Infrastructure found when it reviewed roughly 8,000 flood-control projects and confirmed about one in twenty as anomalous. To be precise about what that comparison is: choosing a cut so the flagged share sits below a known confirmed rate is a conservative budgeting choice, not corroboration. We have not matched individual flagged sites against the Commission's findings, and the two lists should not be assumed to overlap. The cut itself is a fixed parameter (absence score ≥ 0.62, published with the data files and in the open-source bake script) so anyone can re-derive the red set or move the cut and see what changes. A red marker is a prompt for ground-truth review, never proof: narrow or small structures can be genuinely built yet sit below optical resolution.
Historical Imagery for Every Project
The Sentinel-2 verdict drives the marker colour. Every project on the map also opens an on-demand before/after view built from the Esri World Imagery Waybackarchive: high-resolution historical basemap snapshots from 2014 to the present. Pick two dates and drag to compare how a site changed over time. Where construction is not visible, look for whether the structure ever actually appears.
This view is imagery only and carries no automated verdict. The dates are when Esri refreshed its basemap, not necessarily when the area was re-photographed, so some locations look unchanged between two dates, and the viewer says so when that happens. It is a way to look with your own eyes, kept deliberately separate from the Sentinel-2 change detection so raw imagery is never mistaken for an accusation.
Composite Strategy
Rather than a single image, each site gets a median composite of every cloud-free Sentinel-2 scene in a before window (the year ahead of the project's funding year) and an after window (the most recent two years). The median strips out clouds, shadows, and passing features. Where neither window has a clear scene, the project is left unassessed rather than guessed.
Known Limitations
Narrow or water-adjacent structures
A bridge span over a river, or a flood-control wall along a bank, is a thin line of concrete inside a buffer that is mostly water and bare ground. Averaged over the area, the built-up signal can stay below threshold even when the structure was genuinely completed. This is a common reason a real, finished project shows little visible change, and the main reason most flood-control sites read as weak signal. Treat such reads as a prompt to look closer, not a finding.
Underground infrastructure
Pipes, cables, and underground utilities produce no surface change. These will show as No Change regardless of actual completion.
Small-footprint projects
Structures smaller than ~500m² are below Sentinel-2's 10m resolution and may not be detectable.
Dense urban replacements
Demolishing and rebuilding in an already built-up area may produce little net NDBI change.
Cloud-heavy regions
In tropical regions with 200+ cloudy days a year, a site may have no clear before or after scene in the imagery window. When neither window has a usable composite, the project is left unassessed rather than guessed.
Climate zone calibration
Default thresholds are calibrated for tropical Philippines. Arid or temperate zones require threshold recalibration against known-good projects.
Disclaimer
Satellite reads are automated change-detection on free 10m Sentinel-2 imagery and can be wrong: small or narrow structures, projects completed outside the imagery window, and persistent cloud cover are common reasons a genuinely built project shows little visible change. A flagged project is a prompt for review, never proof of wrongdoing. Every case needs ground-truth investigation before any conclusion is drawn. All figures are from the public DPWH record.