Why study repeated oil-spill reports?
What can the records tell us?

About the author

The work demonstrates how mapped records and transparent model comparisons can support historical investigations without overstating what the available reports can establish.

Purpose and scientific position

We examined accepted spill reports to understand where documented incidents occurred more than once, and to identify places whose recorded histories merit closer examination. Historical recurrence can support investigation and further monitoring design, but it is not a forecast of future spills, a measurement of contaminated land or water, or evidence of responsibility.

Why the study covers 2006–2018

This is a retrospective study, not a current spill-monitoring service. We used a frozen historical collection of documented oil-spill records from the NOSDRA evidence system, whose earlier nationwide material was also published through UNEP. The UNEP catalogue describes a historical 1994–2019 source; our selected analysis window is narrower and should not be confused with the full coverage of that source.

We began in 2006, the year NOSDRA was established and the starting point used for the agency's modern published reporting period. For the historical analyses we selected records through 2018: earlier years provided historical evidence, while 2018 observations were kept separate for a later-year evaluation of patterns established from reports available up to 2017.

Why not 2019–2026? Newer records were not acquired, reconciled and scientifically validated within this locked atlas. Their absence here does not mean that no spills occurred after 2018 or that newer information is unavailable. NOSDRA maintains an ongoing Oil Spill Monitor. Extending this atlas would require a separate update using clearly dated source records, incident and location quality checks, consistent study-cell assignments, and revised temporal evaluation.

Limits of the historical evidence: The reporting record can also be incomplete within 2006–2018. A year included in the study must not be described as an exhaustive record of every physical spill. These results help examine documented recurrence, not current spill activity, direct exposure, environmental damage or future incident probability.

Validated source counts

• 3,109 documented events in 2006–2018 on 107,133 model cells (nominal 300 m).
• 3,078 of these events are pre-2018; 2018 has 31 recorded events in 29 positive cells.
• The 2018 recurrence classes use ONLY 2017-and-earlier incidents. Observed 2018 labels appear exclusively in independent evaluation summaries.

Descriptive 2018 recurrence evidence

• Cells with at least one recorded event during 2015–2017: 312, representing 0.291% of model cells.
• These cells contained 14 of 31 documented 2018 events (45.2%). This does not demonstrate reliable performance at new, previously unrecorded locations.

Machine-learning comparison and limitations

• Rolling 2016–2018 holdout AP: logistic ML 0.189978; simple recent-three-year count 0.241718.
• The simple recurrence benchmark outperformed ML. Novel-location discrimination was weak. The scores must not be interpreted as calibrated physical spill probabilities.
• Validation was rolling annual using strictly earlier history, rather than a three-year 2015-frozen forecast. Discrete count ties make exact top-k screening thresholds ambiguous.

Contextual population, built area, and historic water

• WorldPop 2018 modelled population: approximately 2,040,673 in the source-supported allocated cells; 104,718/107,133 cells supported.
• GHSL 2015 reconstructed built surface: 38,279,288 m² approximate allocated total; 106,838/107,133 cells supported.
• WorldPop and GHSL estimates use native-grid pixel-centre allocation and have boundary and modelling uncertainty. These are NOT numbers of people or structures affected by spills.
• GHSL 2015 represents an R2023A historical reconstruction, not a contemporaneous published 2015 layer. JRC 2015–2017 surface-water occurrence does not reveal oil pollution.

Cartographic changes at R3

• Five A3 landscape maps (PNG and PDF). Occurrence and recurrence maps use scaled point symbols at computational-cell centres and two count-density-selected insets. Marker sizes are symbolic, not physical spill extent.
• Population and built-area maps use log-scaled positive-value displays capped at the 99th percentile for visualization only. Underlying data remain unchanged. Missing coverage is separate from a valid zero.
• Include UTM coordinate frame, north arrow, scale bar, scientific disclaimer, and map-specific attribution. Insets are labelled A and B without asserting unverified named settlements.

Data and attribution

• Accepted NOSDRA documented spill records (2006–2018); accepted HDX Bayelsa boundary; 300 m model grid.
• Satellite and ancillary sources from locked earlier processing: JRC Global Surface Water, Landsat, CHIRPS, NASADEM.
• WorldPop GP 2018; JRC GHSL GHS-BUILT-S R2023A (2015 reconstructed epoch).
• Model and audit: G10C.2, G11, G11B. This R3 reads checksum-verified G12 R2 results and does not retrain or recompute the scientific baseline.

Scientific restrictions

• Unknown/unreported spills are not observable zeros. No confirmed spatial contaminated extent, remediation cost, direct damage or health exposure is mapped.
• No oil-infrastructure presence/absence model can be claimed. Infrastructure coverage was deemed insufficient for historical backtesting.
• Do not label maps as oil-spill probability or confirmed environmental contamination. Descriptive screening should guide field verification, not substitute for it.
• The public interactive website is hosted on GitHub Pages. Its map library and OpenFreeMap basemap require an internet connection. Successful website loading does not establish field verification or the accuracy of unreported incident locations.

Publication & preservation

The evidence atlas is published on GitHub Pages. The original G12 R3 scientific results and cartography passed their documented review. Later website changes clarify presentation without altering historical classifications, mapped evidence locations or scientific conclusions.

The repository's release_manifest.json records checksums for the original G12R4.2 release; it is retained as an archival record rather than recalculated when editorial files change. Source cartography: G12R3_FINAL_CARTOGRAPHY_20261009.