How BlackTide works

The Niger Delta's 50,000 km² of creeks, mangroves and swamp make ground surveys slow, dangerous and incomplete, and official spill figures vary widely between agencies and operators. BlackTide watches the whole region from orbit, every month, and makes what it finds open to everyone.

1. 01

   Sea & coast: SkyTruth Cerulean

   Oil on open water is taken from SkyTruth Cerulean, which runs a deep-learning slick detector on every Sentinel-1 radar pass and has part of its output reviewed by people. BlackTide adds context to each slick: nearest community, people within 5 km (WorldPop) and mangrove area.

2. 02

   Land, swamp & creeks: official reports

   Every spill reported to NOSDRA since 2015 (Nigerian Oil Spill Monitor, from Joint Investigation Visits) is mapped with its operator, reported cause, volume and habitat. For land and swamp reports, BlackTide checks from space whether the spill left visible damage: vegetation change at the site (100 m) against its surroundings (300–1,000 m), dry season before vs after, on Sentinel-2 NDVI and Sentinel-1 radar. Tested on 133 reports vs 139 spill-free points: clear local die-back at 23% of reported spills and 1% of controls.

3. 03

   Land, creeks & estuaries: BlackTide detection (in development)

   Cerulean doesn't cover inland areas, so BlackTide looks for what oil does there: it kills vegetation. Dry-season Sentinel-2 NDVI is compared year on year to find sudden die-off of healthy mangrove, swamp forest and farmland, flagging patches beside creeks.

4. 04

   Human review

   Die-off has other causes too (clearing, fire, flooding), so each candidate is checked by an analyst on before/after imagery. Confirmed and rejected sites both become training labels.

5. 05

   Land model

   A Random Forest trained on those labels uses monthly Sentinel-1 radar (which sees through cloud) and Sentinel-2 indices — NDVI change, NDWI, MNDWI, the Oil Spill Index — with a radar-only fallback when clouds hide every optical scene.

6. 06

   Research comparison

   BlackTide's own marine models (Random Forest and a neural network on per-pass radar, local darkness, wind and incidence angle) are trained on Cerulean's reviewed slicks and reported on the Model page — but not used for the map, because pixel-level models raise too many false alarms over open sea.

7. 07

   Validate & repeat

   Everything runs on free tools. A monthly GitHub Action imports new Cerulean slicks, runs the land model and republishes the site.

Known limitations

• —Dark patches on radar are not always oil: low wind, rain cells, algae and mudflats look similar. The model learns these look-alikes only if they are labelled.
• —Land spills under dense mangrove canopy are harder to see than slicks on open water.
• —Sentinel-1B failed in December 2021, halving revisit frequency until Sentinel-1C came online in 2025 — fewer looks means more missed short-lived slicks in that period.
• —Land events are dated by the month (or dry season) they were detected, not the exact spill date; marine events carry the exact satellite pass date.
• —Cerulean's machine detections that no one has reviewed yet are shown as unverified.

Tools built on

Research team

ThankGod Chinemerem Ugwuada · Ikenna Okonkwo Anthony · Tochukwu Ambrose Ngwu

Marine and coastal slick detections: SkyTruth Cerulean (cerulean.skytruth.org). Official spill reports: National Oil Spill Detection and Response Agency (NOSDRA), Nigerian Oil Spill Monitor (oilspillmonitor.ng). Boundaries: geoBoundaries (CC BY 4.0). Contains modified Copernicus Sentinel data. Sentinel-2 cloudless mosaics © EOX IT Services GmbH, CC BY-NC-SA 4.0.