xquantize/global-natural-events-eonet-2011-2026

Dataset Card for Global Natural Events (NASA EONET, 2011–2026)

Overview

This dataset contains a clean, analysis-ready record of 41,600 natural-hazard observations spanning August 30, 2011, to July 23, 2026. Built directly from NASA's Earth Observatory Natural Event Tracker (EONET), the data has been flattened, validated, and enriched with location and unit-normalization fields that the raw API does not natively provide.

• Total Records: 41,600 observations across 19,281 unique events
• Temporal Range: 2011-08-30 to 2026-07-23
• Hazard Representation: 13 distinct natural hazard types (dominated by Severe Storms and Wildfires)
• Source: NASA EONET API (Aggregated from USGS, NOAA/NHC, JTWC, GDACS, InciWeb, NASA Earth Observatory, and others)

---

What Makes This Dataset Different?

The raw EONET API is notoriously challenging to work with: geometry is deeply nested, a single event spans multiple observations, coordinates are in GeoJSON [lon, lat] order, magnitudes mix incompatible units within a single field, and area events arrive as polygons with no point coordinate.

This dataset resolves those structural issues:

1. Flattened Observation Rows: One row per observation. Multi-point events (storm tracks, long-burning fires) span multiple rows. You can group by event_id to collapse back to the event level.
2. Location Enrichment: Country, ISO code, admin region, and nearest populated place added for 41,584 of 41,600 rows via offline reverse-geocoding.
3. Honest Magnitude Handling: Raw values are kept intact; a magnitude_kind column clarifies what each number measures, and fire_area_hectares normalizes mixed acre/hectare fire reporting into a single metric.
4. Confidence Flags:
   • likely_offshore: Identifies ocean points whose "nearest land" geocoding is approximate.
   • coord_is_centroid: Flags the 297 polygon footprints reduced to a representative center point.

---

Important Modeling & Analysis Notes

• Incomparable Magnitudes: Never aggregate magnitude_value across hazards, as it mixes wind speed (knots), burned area (hectares/acres), and ice area (square nautical miles). Always segment by magnitude_kind.
• Structural Nulls: ~45% of rows have no magnitude. This is structural (many hazard types simply do not report magnitudes), not missing data.
• Class Imbalance: Severe Storms (47%) and Wildfires (40%) account for ~87% of all rows. Rare hazards have too few rows for standalone modeling and are best treated as a combined "other" class or used descriptively.
• Catalog Scope: EONET is a curated catalog of notable events, not an exhaustive census. Absence of an event does not imply nothing occurred.

---

Known Limitations & Edge Cases

• Missing Coordinates (16 rows, 0.04%): 5 rows are Points whose raw coordinates fell outside valid lat/lon ranges and were nulled. 11 rows are Polygon centroids that computed to invalid coordinates (likely footprints straddling the antimeridian at $\pm 180^\circ$ longitude). These rows retain event, hazard, date, and magnitude info but are not mappable.
• Polygon Centroids (297 rows): Represented by their footprint centroid (coord_is_centroid = True). Suitable for regional or macro-scale analysis, but filter these out for point-precision work.
• Offshore Geocoding: Open-ocean points map to the nearest populated place (often tiny remote islands like South Georgia or Cocos Islands), inflating observation counts for those territories. Use likely_offshore to filter them if necessary.

---

Schema Reference (Core Columns)

Column Name	Data Type	Description
event_id	String	Stable EONET identifier (e.g., EONET_21399)
event_title	String	Human-readable event name (e.g., "Hurricane Fausto")
hazard_id	String	Machine-readable slug (severeStorms, wildfires)
hazard_type	String	Human-readable category (13 distinct values)
obs_date	DateTime (UTC)	Timestamp of this specific observation point
is_closed	Boolean	True if closed/ended, False if ongoing
latitude	Float	Latitude in decimal degrees ($-90 \text{ to } 90$)
longitude	Float	Longitude in decimal degrees ($-180 \text{ to } 180$)

(Note: The full CSV contains 24 enriched columns including administrative regions, magnitude types, and confidence flags).

---