HeimdallResearch/storm-blackout

Storm Blackout and Restoration

Not for real-time emergency response or for any decision about one address or property.

Not in this download, free through the API: lights-out videos for most storms, storm-path videos for hurricanes and tropical storms, forecast maps and satellite loops. Get a free key at https://heimdallresearch.com/go/signup

Night light data: NASA Black Marble (VNP46A2, VJ146A2), courtesy of NASA LAADS DAAC. Roman et al. (2018), DOI 10.1016/j.rse.2018.03.017.

Utility outage data: EAGLE-I / ORNL, Figshare DOI 10.6084/m9.figshare.24237376 (CC BY 4.0).

Customer estimates: U.S. Energy Information Administration Form EIA-861 and U.S. Census Bureau county housing unit estimates (public domain).

Declarations, boundaries and tracks: OpenFEMA, U.S. Census Bureau cartographic boundary files and NOAA NHC best tracks (public domain).

Not for real-time emergency response or for any decision about one address or property. Provided as is, without warranty of any kind.

County records for 672 FEMA-declared U.S. storms with landfall or incident dates from 2012-01-14 to 2026-08-11: NASA Black Marble night light by county and night, and reported power outages and restoration times from DOE EAGLE-I utility records where they exist.

The catalog lists 672 storms in storms.csv. 658 have night-light numbers: 611 by county and 47 by tribal area. 477 have an outage run on EAGLE-I, and 464 of these have records for at least one county. Every storm without numbers keeps its row, with a reason code in night_light_reason or outage_reason.

Night light is the county mean of NASA Black Marble daily radiance over valid pixels, one night per row in county_daily.csv, with a per-county summary in county_summary.csv. Each storm is measured with one product, named in product. Radiance from different products should not be compared between storms; this release gives no conversion between them.

The outage columns come from EAGLE-I county customers out (Figshare DOI 10.6084/m9.figshare.24237376) for the declared counties of each storm, in a 45-day window from each county's impact date.

47 storms were declared for a tribal area with no county named. They ship as tribal-area rows (tribal_area_summary.csv, tribal_area_daily.csv) and county context rows (region_role is county_context). County rows for tribal-area storms are context: they describe the whole county, not the tribal area, and tribal nations are sovereign areas. The share fields say how much of each county lies inside the area.

An empty cell means the value was not measured. It is never written as 0. A column that can be empty for more than one reason has a reason column beside it, and docs/DATA_DICTIONARY.md says for every column when it is empty and why.

Key figures

• Storms in the catalog: 672
• Storms with night-light numbers: 658
• Storms with an outage run: 477
• Counties and county equivalents with night light: 3,194
• Storm-county rows with utility records in the window: 10,260
• Median peak customers out over storm-county rows with reason ok: 1,766
• Median hours to 90% restored, storm-county rows where it was measured: 30.2 hours

What the data shows

Irma has the largest sum of county outage peaks among these 672 storm events: 8.9 million.

Among these 672 storm events, Irma (AL112017) has the largest sum of county peak customers out: 8,867,374 across 329 counties with a numeric peak. Each county contributes its own peak. Those peaks are not one simultaneous count of customers out.

If you are comparing these storm events, this shows the scale of Irma's county peaks. It is not an outage feed for one day and not a figure for one address.

import pandas as pd
u = pd.read_csv("data/utility_by_county.csv")
x = u.assign(p=pd.to_numeric(u.peak_customers_out, errors="coerce")).dropna(subset=["p"])
g = x.groupby("event_id").p.agg(["sum","count"]).sort_values("sum", ascending=False)
print(len(pd.read_csv("data/storms.csv")), g.index[0], int(g.iloc[0]["sum"]), int(g.iloc[0]["count"]))

Milton's sum of county outage peaks ranks fifth: 3.7 million.

Milton (AL142024) sums to 3,736,269 peak customers across 51 counties (5th largest sum). Each county contributes its own peak.

Milton is a Florida landfall in this package. Compare its summed peaks with the other storm events in this table.

import pandas as pd
u = pd.read_csv("data/utility_by_county.csv")
x = u.assign(p=pd.to_numeric(u.peak_customers_out, errors="coerce")).dropna(subset=["p"])
g = x.groupby("event_id").p.agg(["sum","count"]).sort_values("sum", ascending=False)
print(int(g.loc["AL142024","sum"]), int(g.loc["AL142024","count"]), list(g.index).index("AL142024")+1)

After Maria, night-light brightness took 121 days to return to the usual range in the typical county.

For Maria (AL152017), the median of days_to_normal is 121 days across 46 counties with a value. This column is satellite night-light brightness.

Grid restoration hours and satellite night-light recovery answer different questions on the same storm. Maria's night-light timing sits in Puerto Rico and USVI context in this package.

import pandas as pd
c = pd.read_csv("data/county_summary.csv")
d = pd.to_numeric(c[c.event_id=="AL152017"].days_to_normal, errors="coerce").dropna()
print(int(d.median()), int(d.size))

After Irma, the typical county reached 90% of customers restored in 53 hours.

For Irma, the median county hours to 90 percent restored is 52.9 hours (280 counties with a numeric value).

If you need both clocks on Irma, read this next to the Maria night-light finding. Utility restoration timing and night-light timing are not the same measure.

import pandas as pd
u = pd.read_csv("data/utility_by_county.csv")
h = pd.to_numeric(u[u.event_id=="AL112017"].hours_to_90pct_restored, errors="coerce").dropna()
print(round(float(h.median()),1), int(h.size))

44 of 672 storm events are tropical cyclones or typhoons.

These 672 storm events include 36 with kind tropical and 8 with kind typhoon (44 together). Kind tropical includes hurricanes. Other counts: severe storms 347, floods 186, tornadoes 24, ice storms 21, winter storms 21, snowstorms 20, straight-line wind 6, coastal storms 3.

If you expected only hurricanes, check this first. Severe storms and floods are the two largest counts. Kind tropical includes hurricanes.

import pandas as pd
s = pd.read_csv("data/storms.csv")
u = pd.read_csv("data/utility_by_county.csv")
x = u.assign(p=pd.to_numeric(u.peak_customers_out, errors="coerce")).dropna(subset=["p"])
print(len(s), s.kind.isin(["tropical", "typhoon"]).sum(), s.kind.value_counts().to_dict(), list(x.groupby("event_id").p.sum().nlargest(3).index))

Coverage

Publisher	Heimdall Research
Version	1.2
Released	2026-10-10
Version DOI	10.5281/zenodo.23010005
Concept DOI	10.5281/zenodo.23010004 (all versions)
License	CC BY 4.0
Web page	https://heimdallresearch.com/data/storm-blackout/
Period	2012-01-14 to 2026-08-11 (672 storms)
Resolution	County by day, and tribal area by day for tribal declarations
Update cadence	This catalog runs through 2026-08-11. New storms are not added on a schedule.

Versions

Version	Where published
1.0	Kaggle
1.1	Kaggle
1.2	Zenodo, Kaggle, Hugging Face (first version on Zenodo and Hugging Face)

Versions 1.0 and 1.1 were published on Kaggle only and have no Zenodo DOI. Version 1.2 is the first version on Zenodo and Hugging Face.

Spatial units:

area	period	storms	counties	tribal_areas
United States, Puerto Rico and U.S. Virgin Islands, with Guam, the Northern Mariana Islands and American Samoa as storm rows only	2012-01-14 to 2026-08-11	672	3,194	28

Research use

Questions the tables can answer, each with the table and columns it uses, and what the tables leave open: docs/RESEARCH_USE.md.

Method

• Storms: storms.csv lists FEMA-declared hurricanes, tropical storms, typhoons, winter storms, severe storms and floods. storms.landfall is the UTC date of the first HURDAT2 landfall on a U.S. state (the first on a territory when there was none on a state). For FEMA-numbered storms it is the incident begin date. storms.source gives the source of each date. A FEMA disaster that is the same storm as an NHC storm, with the same landfall date, states, declared counties and outage records, has no row of its own: its number is in also_known_as of the NHC storm. The NHC storm's night-light rows are kept, because their impact dates come from the track.
• Declared counties: a county is declared when the storm's major disaster declaration names it. For tropical storms it is also declared when the storm's emergency declaration names it, matched by year and storm name in the declaration title. A title that names two storms counts for both. declared_via says which. A county named in both is major_disaster.
• Night light: NASA Black Marble daily radiance (VNP46A2 on Suomi NPP, VJ146A2 on NOAA-20), dataset DNB_BRDF-Corrected_NTL. The gap-filled dataset is not read. A pixel is valid when it has a value, Mandatory_Quality_Flag is 0, the cloud mask says confident clear and Snow_Flag is 0. County radiance for a night is the sum of radiance over the county's valid pixels, from every tile it touches, divided by the number of those pixels. It is empty when there is no valid pixel.
• One product per storm: the baseline and the nights after the storm come from the same satellite, and product names it.
• Boundaries: Census cartographic boundary file cb_2024_us_county_500k for counties, cb_2020_us_county_500k for the eight Connecticut counties, and cb_2024_us_aiannh_500k for tribal areas. The file hashes are in census_boundaries_manifest.json.
• Summary: each county has an impact date from the HURDAT2 track (first date within 250 km, or the date of closest approach), or the catalog landfall date. Clear nights are smoothed with a 5-night rolling median and compared with the county's own range before the storm, giving status, days_to_50 and days_to_normal.
• Outage: EAGLE-I county customers out for the declared counties of each storm, in a 45-day window from each county's impact date. The window ends earlier at the next FEMA storm in the county (outage_window_end_reason). Readings that hold more than one collection snapshot are flagged and replaced before peaks, customer-hours and restore hours are taken. Restore times count hours from the peak until customers out stay at or below 50% and 10% of the peak for 6 hours. Outage method v2. A reporting drop is one 15-minute step that cuts the outage by more than 25% of the county peak and lands at or below 10% of the previous reading. When that low spell lasts the 6-hour sustain and the feed does not return to at least half the pre-drop reading within 12 hours, the restore hours that the spell would have set are empty with reason reporting_drop, and reporting_drop is true. Customer-hours stop at that step (customer_hours_reason). If the feed returns within 12 hours, those readings are skipped and the hours are read from the series after the return. No restoration time is filled in. A reading at the start or end of a county series has only the reading beside it. The spike rule looks at both sides, so it leaves that reading in place. When setting it to the reading beside it makes the recomputed peak match the published peak, the restore hours are read from that series. Where the 15-minute series does not reproduce the published peak, the restore hours are empty with reason peak_mismatch and the peak is left as published.
• Customers: EIA-861 customers by state and utility, split across counties by Census housing units, for the EIA year nearest the storm (customers_source).
• Empty values: a value that was not measured is left empty, with a reason column where there can be more than one reason. It is never written as 0.

Intended use: Comparing reported outages and restoration times across counties and storms, and studying how county night light changes after a storm.

Not intended for: Real-time emergency response, any decision about a single address or property, or statements about a tribal area drawn from county rows.

Validation

Check	Result	Source
Release gates	Every rule in the dataset spec holds on every row of every table: filled and unique keys, joins between tables, a reason for every empty value that has more than one reason, one product per storm, thin_night for 1 to 19 valid pixels, declared_via empty exactly when the county is not declared, and radiance empty exactly when there is no valid pixel.	release/datasets/storm_blackout.yaml
Counts in this text	Every count in the README, the changelog and the limits is read from the tables when the package is built. The build stops when one differs from the release plan.	release/checks/storm_blackout_numbers.py
Rows carried from the per-storm files	Every row of every per-storm file in the final cut is in the tables, storm by storm, and the statewide county files of the tribal declarations are not.	pipelines/storm_blackout/test_catalog_tables.py (heimdall-data)
Night light recomputed from the night-light table	20 storms (tropical, winter, severe storm and flood, Connecticut, Puerto Rico, tribal-area and late-2025 storms) were derived again from the night-light table. Every summary and nightly value equals the tables within a relative 1e-9, apart from the values the publication rules leave empty.	pipelines/storm_blackout/test_catalog_tables.py (heimdall-data)
Outage recomputed from the EAGLE-I file	The outage columns of 5 counties (Lee, Charlotte and Orange in Florida for Ian; Ponce and Mayagüez in Puerto Rico for Fiona) were computed again from the published EAGLE-I 2022 file and equal the tables exactly.	pipelines/storm_blackout/test_catalog_tables.py (heimdall-data)
Night light against reported outages	County mean darkness (dark_county_mean) does not separate counties with large outages from counties with small ones: pooled AUC 0.515 over 418 storms and 5,841 counties, 95% interval 0.482 to 0.547 from resampling storms. The interval includes 0.5, which is chance. Lit-pixel radiance drop (dark_lit_ratio) separates counties with large outages from counties with small ones: pooled AUC 0.559 over 287 storms and 3,611 counties, 95% interval 0.511 to 0.605 from resampling storms. Change in the share of unlit lit-mask pixels (dark_lit_share) separates counties with large outages from counties with small ones: pooled AUC 0.566 over 287 storms and 3,611 counties, 95% interval 0.513 to 0.615 from resampling storms. An AUC of 0.5 is a coin flip. Per-storm scores for 477 storms are in validation.csv, pooled scores in validation_pooled.csv, and scores with whole groups of states held out in validation_holdout.csv. FEMA DR-4466 is recorded on AL112019 in storms.also_known_as and has no validation row.	validation.csv, validation_pooled.csv, validation_holdout.csv; pipelines/storm_blackout/county_day/validation.py (heimdall-data)

Limits and appropriate use

• Night light is related to electricity service but is a different quantity. Generators can restore light before the grid does, and damaged or empty buildings can stay dark after power returns. status, days_to_50 and days_to_normal come from night light alone. Scored against reported outages, the result is weak. County mean darkness (dark_county_mean) does not separate counties with large outages from counties with small ones: pooled AUC 0.515 over 418 storms and 5,841 counties, 95% interval 0.482 to 0.547 from resampling storms. The interval includes 0.5, which is chance. Lit-pixel radiance drop (dark_lit_ratio) separates counties with large outages from counties with small ones: pooled AUC 0.559 over 287 storms and 3,611 counties, 95% interval 0.511 to 0.605 from resampling storms. Change in the share of unlit lit-mask pixels (dark_lit_share) separates counties with large outages from counties with small ones: pooled AUC 0.566 over 287 storms and 3,611 counties, 95% interval 0.513 to 0.615 from resampling storms. An AUC of 0.5 is a coin flip. Use the utility columns for outage size and restoration time.
• Radiance from different products should not be compared between storms; this release gives no conversion between them. 500 storms use VNP46A2 and 158 use VJ146A2 (product). Compare radiance within a storm, or between storms on the same product.
• County rows for tribal-area storms are context: they describe the whole county, not the tribal area, and tribal nations are sovereign areas. The share fields say how much of each county lies inside the area. The 47 tribal declarations have no county list, so they have tribal-area rows and overlapping-county context rows only and are absent from county_summary.csv and county_daily.csv. Their outage and customer columns are empty: EAGLE-I reports counties (no_county_outage_record) and EIA-861 customer counts are by county (no_county_customer_count). Outage records for the overlapping counties of 41 of these storms are in tribal_context_utility_by_county.csv, as context.
• Henri (AL082021) lists MA in states because of FEMA EM-3566-MA, an emergency declaration for the Mashpee Wampanoag Tribe that names no county. Its 14 Massachusetts rows in county_summary.csv are measured but not declared (reason is county_not_declared), and the tribal area, which lies in Barnstable and Bristol counties, is not measured for this storm.
• thin_region is True for a tribal area with fewer than 20 pixels on its tiles. thin_night is True for a county or area night with 1 to 19 valid pixels. Both keep the value as measured, and such values are noisy.
• 14 storms have no night-light numbers: 3 have a storm window that starts before the VNP46A2 record begins on 2012-01-19 (before_nasa_record), and 11 are in Guam, the Northern Mariana Islands or American Samoa, which the county boundary file does not cover (territory_not_in_county_boundaries).
• 195 storms have no outage run: 117 predate the public EAGLE-I county series, which starts 2014-11-01 (no_utility_record_before_2014). 30 fall in a year with no published EAGLE-I file (no_published_year). 41 are tribal declarations with no county named (no_county_outage_record). 7 are in territories EAGLE-I does not cover (territory_not_in_eaglei).
• 13 storms have an outage run but no county with EAGLE-I records in the window, so none of their utility_by_county.csv rows has reason ok: AL152017, DR-4227, DR-4271, DR-4275, DR-4296, DR-4306, DR-4311, DR-4327, DR-4388, DR-4571, DR-4609, DR-4649, DR-4805. Their rows give no_utility_record or no_rows_in_window, and their outage columns are empty.
• The 6 late-2025 storms (DR-4901, DR-4902, DR-4903, DR-4905, DR-4906, DR-4907) have an outage window that ends 31 December 2025, the last day of the published EAGLE-I record (window_note is window_cut_at_source_end). outage_window_end_day gives the planned window for these storms, and their records stop earlier.
• Customer-hours for counties in both version 1.1 and this table: Matthew (AL142016) sums to 140,943,783.1 in version 1.1 and 140,435,077.7 here (70 of 181 counties differ); Irma (AL112017) sums to 595,686,772.4 in version 1.1 and 586,489,025.7 here (144 of 379 counties differ).
• County pixels are placed half a pixel off: the first longitude and latitude of each tile are read as a pixel centre, where the file gives the tile's west and north edge. This release does not measure how many pixels land in a different county or how much county mean radiance would change. The numbers are kept as they are.
• Connecticut uses the eight 2020 counties; the 2024 file has nine planning regions. The storm records, the declared lists and the customer table use the old counties.
• Hurricane Eta stays as two rows. AL292020 is the NHC storm, and DR-4588 is FEMA's North Carolina disaster declaration titled Tropical Storm Eta. Every county declared for DR-4588 is also declared for AL292020. In those counties AL292020 dates the impact from the track, one day after the FEMA incident begin date that DR-4588 uses, so their night-light and outage values for those counties differ. storms.csv therefore has one row more than it has distinct storms. Count Eta once when counting storms.
• Bedford city, Virginia (51515) is declared but is not in the boundary file; it is listed in declared_unmatched. That table holds 4 rows, all for 51515: AL062018, AL182012, DR-4072, DR-4644. Their outage rows in utility_by_county.csv have an empty county.
• in_nhc_wind_radius is filled for the 34 storms with a HURDAT2 34-knot wind radius on file. Empty means no radius is on file for the storm. It does not mean the county was outside the radius.
• EAGLE-I coverage is incomplete in early years and in some states, and customers_total is an estimate: each state's EIA-861 customer count split across counties by Census housing units. Where the peak is above the estimate the peak is published as reported, and the share is empty with a reason.
• Where the 15-minute series does not reproduce the published peak, the restore hours are empty with reason peak_mismatch and the peak is left as published.
• The word outage still names night-light states: the status value no_measurable_outage, the columns outage_min_ratio and outage_min_day, and the reason code no_outage_measured. They describe the night-light series and say nothing about reported outages.
• valid_pixel_share and snow_excluded_share have different bases (all county pixels, and pixels with a finite radiance), so the two can add to more than 1.

Imagery

Before-and-after night-light pictures are not in this release.

Files and fields

File	Description	Bytes	SHA-256
data/storms.csv (CSV)	One row per storm, including the storms that have no night-light numbers or no outage run	151569	7869fd3f500efa21...
data/storms.parquet (Parquet)	Same rows, columnar format	26268
data/county_summary.csv (CSV)	One row per storm and county with night light (method v3), with the outage columns joined from utility_by_county	19794023	05cfaadf0babcffd...
data/county_summary.parquet (Parquet)	Same rows, columnar format	2099754
data/county_daily.csv (CSV)	One row per storm, county and night	3050422877	1c508796720c2c72...
data/county_daily.parquet (Parquet)	Same rows, columnar format	707570364
data/utility_by_county.csv (CSV)	Outage columns, one row per storm with an outage run and declared county	2216098	b2135c8d659f9fe5...
data/utility_by_county.parquet (Parquet)	Same rows, columnar format	416483
data/declared_unmatched.csv (CSV)	Declared counties with no row in the boundary file, kept with a reason	740	35d865cd7c7d20fe...
data/declared_unmatched.parquet (Parquet)	Same rows, columnar format	2181
data/tribal_area_summary.csv (CSV)	One row per tribal declaration storm and tribal area, the area FEMA named	17113	61e2fc5eb0d5ccc2...
data/tribal_area_summary.parquet (Parquet)	Same rows, columnar format	14623
data/tribal_area_daily.csv (CSV)	One row per tribal declaration storm, tribal area and night	3347545	6de49dbfa55b71a7...
data/tribal_area_daily.parquet (Parquet)	Same rows, columnar format	823762
data/tribal_area_counties.csv (CSV)	The counties each tribal area overlaps, with the overlap as a share of each. Slivers are left out	7565	fc2437a3053be82e...
data/tribal_area_counties.parquet (Parquet)	Same rows, columnar format	8475
data/tribal_context_county_summary.csv (CSV)	Night-light summary of each county a tribal area overlaps, for the tribal declaration storms. Context rows for the whole county	43881	33c410315abb5e4e...
data/tribal_context_county_summary.parquet (Parquet)	Same rows, columnar format	20560
data/tribal_context_county_daily.csv (CSV)	One row per tribal declaration storm, overlapping county and night. Context rows for the whole county	11276567	3f878b1154efbbd6...
data/tribal_context_county_daily.parquet (Parquet)	Same rows, columnar format	3076213
data/tribal_context_utility_by_county.csv (CSV)	Outage columns of each county a tribal area overlaps, for the tribal declaration storms. Context rows for the whole county	32155	3ff741c11d86f961...
data/tribal_context_utility_by_county.parquet (Parquet)	Same rows, columnar format	18794
data/validation.csv (CSV)	Night light against reported outages, one row per storm and night-light measure, for the storms with an outage run	160389	b9f51e93b2f45982...
data/validation.parquet (Parquet)	Same rows, columnar format	18936
data/validation_pooled.csv (CSV)	Night light against reported outages pooled over storms, one row per measure and scope	6402	f4ab57337e90d386...
data/validation_pooled.parquet (Parquet)	Same rows, columnar format	5489
data/validation_holdout.csv (CSV)	Pooled AUC on each fold of a grouped holdout that keeps whole groups of states together	1100	110ee8f1abbf4b67...
data/validation_holdout.parquet (Parquet)	Same rows, columnar format	2139
docs/RESEARCH_USE.md	Research-use notes built from the tables	12256	cf96342b309ea8ee...
docs/questions/q1.md	Research-use notes built from the tables	4585	ae2c853b39fed10b...
docs/questions/q2.md	Research-use notes built from the tables	8221	34687eb82074c110...
docs/questions/q3.md	Research-use notes built from the tables	5112	f80bf24a8df830f7...
docs/questions/q4.md	Research-use notes built from the tables	3084	aff8029461c39c05...
docs/questions/q5.md	Research-use notes built from the tables	3357	bc9ba8eb6b395274...
docs/questions/q6.md	Research-use notes built from the tables	3383	f2f2745f977f5184...
docs/questions/q7.md	Research-use notes built from the tables	2202	51b1f4657e8310da...
docs/questions/q8.md	Research-use notes built from the tables	1830	e92d5e1a7481bf6f...
docs/questions/irma-largest-summed-county-outage-peak.md	Research-use notes built from the tables	1595	2d216bf1f583a748...
docs/questions/milton-summed-county-outage-peak.md	Research-use notes built from the tables	1318	7306e2ccbc908b5d...
docs/questions/maria-median-days-to-normal-night-light.md	Research-use notes built from the tables	1305	ed7e0c01d2734dbc...
docs/questions/irma-median-hours-to-90-percent-restored.md	Research-use notes built from the tables	1203	2d6cfdf61a40b1d3...
docs/questions/storms-1-2-by-kind-and-top-peaks.md	Research-use notes built from the tables	1434	ec397543c69ee6c8...

On Zenodo downloads, nested paths are flattened: data/foo.csv appears as data__foo.csv. On Kaggle the CSV files sit at the top level: foo.csv.

Also in this package: CHANGELOG.md, LIMITS.md, docs/DATA_DICTIONARY.md.

Fields

storms

One row per storm, including the storms that have no night-light numbers or no outage run

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
name	string		Storm name or FEMA declaration title
kind	string		Storm type from the NHC record or the FEMA incident type
landfall	date		UTC date of the first HURDAT2 landfall on a U.S. state (the first on a territory when there was none on a state). For FEMA-numbered storms, the FEMA incident begin date
landfall_time_utc	string		UTC time of that landfall, HH:MM
first_impact_date	date		UTC date the storm centre first came within 250 km of a county in the table, when that is two or more days before landfall. The baseline nights end before it
states	string		Declared states and territories, space separated
source	string		Where the landfall date and time come from
night_light_unit	string		county: night light by county. tribal_area: night light by tribal area, for a tribal declaration with no county named
night_light_reason	string		Why the storm has no night-light numbers. before_nasa_record: the storm window starts before the VNP46A2 record begins on 2012-01-19. territory_not_in_county_boundaries: Guam, the Northern Mariana Islands or American Samoa, which the county boundary file does not cover
product	string		Night-light product the storm is measured with: VNP46A2 (Suomi NPP) or VJ146A2 (NOAA-20). One product per storm. Radiance from different products should not be compared between storms
outage_reason	string		Why the storm has no county outage run. no_utility_record_before_2014: before the public EAGLE-I county series starts on 2014-11-01. no_published_year: no EAGLE-I file is published for the year. no_county_outage_record: a tribal declaration that names no county. territory_not_in_eaglei: a territory EAGLE-I does not cover
window_note	string		window_cut_at_source_end: the 45-day outage window runs past 31 December 2025, the last day of the published EAGLE-I record, so records stop there
also_known_as	string		FEMA disaster numbers for the same storm that have no row of their own, space separated (for example DR-4466 on Imelda, AL112019). Each such declaration names the same counties on the same date, and its outage records are the same

county_summary

One row per storm and county with night light (method v3), with the outage columns joined from utility_by_county

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
county	string		County name
state	string		State or territory postal code
declared	boolean		True when FEMA declared the county for this storm
declared_via	string		The declaration that names the county: major_disaster, or emergency for a tropical storm's emergency declaration. A county named in both is major_disaster
in_nhc_wind_radius	boolean		True when the county lies inside the NHC 34-knot wind radius of the storm, False when it lies outside
product	string		Night-light product the storm is measured with: VNP46A2 (Suomi NPP) or VJ146A2 (NOAA-20). One product per storm. Radiance from different products should not be compared between storms
baseline_radiance	number	nW/cm²/sr	Median radiance of the clear baseline nights before the storm
baseline_nights	integer		Clear nights in the baseline
baseline_source	string		event_window: baseline nights before the storm. prior_year: the same season a year earlier, used when the nights before the storm were too few. insufficient: no baseline
post_obs_nights	integer		Nights with usable pixels after the landfall date
median_valid_pixel_share	number		Median share of usable pixels after the landfall date
status	string		What the night-light series did. It is not a detected power outage. light_drop: smoothed light fell below the county's normal range soon after impact. no_measurable_outage: observed, no such fall (the name keeps an old use of the word outage for a night-light state). cloud_obscured: not observed until late and normal by then. low_signal: too noisy to judge. insufficient_data: too few nights
outage_min_ratio	number		Lowest smoothed light ratio after impact (a night-light value despite the name)
outage_min_day	integer	days	Day of that lowest ratio, counted from impact_date
days_to_50	integer	days	Days from impact until smoothed light was back at or above 50% of baseline
days_to_50_reason	string		Why days_to_50 is empty. never_below_50: a light_drop status, but light never fell below 50%. not_back_to_50_in_window: fell below 50% and did not return. no_outage_measured: observed, no light drop found (a night-light state despite the name). not_observed_cloud, not_observed_too_few_nights: the county was not observed. not_measurable_low_signal: its light is too noisy to judge
days_to_normal	integer	days	Days from impact until smoothed light was back in the county's normal range
days_to_normal_reason	string		Why days_to_normal is empty. not_back_to_normal_in_window: a light_drop status that did not return to normal. The other codes are as in days_to_50_reason
recovery_uncertain	boolean		True when a gap of 5 or more days without a clear night comes just before days_to_normal
window_end_day	integer	days	Last day of the night-light window, counted from impact_date
window_end_reason	string		window: the full window. data_end: the stored nights stop earlier. next_storm:DR-number: the next FEMA storm in the county
impact_date	date		Date the storm reached the county, day 0 of the window
impact_source	string		How impact_date was chosen. track_within_250km: first UTC date the HURDAT2 track came within 250 km of the county. track_closest_approach: the track never came that close, so the date of closest approach. catalog_landfall: storms.landfall
clear_nights_pre	integer		Clear nights before impact
clear_nights_post	integer		Clear nights after impact
method_version	string		Night-light method version
source_night_light	string		Layer marker
post_days_window	integer	days	Days after the landfall date covered by stored nights
available	boolean		True when utility records cover this county's window
reason	string		Utility availability. ok: records in the window. no_rows_in_window: rows that year but none in the window. no_utility_record: no EAGLE-I rows for the county that year. county_not_declared: the outage run reads declared counties only. no_utility_record_before_2014, no_published_year: the storm-level outage_reason, for a storm with no outage run
source	string		utility_records when the county has EAGLE-I rows in the window, otherwise none
peak_customers_out	number		Peak customers out reported in EAGLE-I, taken after readings that hold more than one collection snapshot are replaced (see flagged_readings). A replaced reading is an interpolated value that was not measured
peak_reason	string		Why peak_customers_out is empty: no_utility_record. reason gives the cause (for example county_not_declared)
peak_customers_out_raw	number		Highest reading in the window as the EAGLE-I year file records it, before any filter
flagged_readings	integer		Readings in the window that held more than one collection snapshot and were replaced. 0 when none
peak_share_out	number		peak_customers_out divided by customers_total
peak_share_reason	string		Why peak_share_out is empty. no_utility_record: no EAGLE-I rows. no_customers_total: customers_total is empty and customers_reason says why. peak_above_customer_estimate: the peak is above the customer estimate
customer_hours_out	number	customer-hours	Customer-hours out in the window from impact_date, or up to a reporting drop when customer_hours_reason is reporting_drop
hours_to_50pct_restored	number	hours	Hours from the peak until customers out stayed at or below 50% of the peak for 6 hours
hours_to_50pct_reason	string		Why hours_to_50pct_restored is empty. not_restored_in_window: still above half the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_to_90pct_restored	number	hours	Hours from the peak until customers out stayed at or below 10% of the peak for 6 hours
hours_to_90pct_reason	string		Why hours_to_90pct_restored is empty. not_restored_in_window: still above 10% of the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_under_1pct_customers	number	hours	Hours from the peak until customers out stayed at or below 1% of customers_total for 6 hours
hours_under_1pct_reason	string		Why hours_under_1pct_customers is empty. never_above_1pct: the outage never passed 1% of customers. not_under_1pct_in_window: still above 1% at the end. no_customers_total, peak_above_customer_estimate or no_utility_record: nothing to measure against. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
coverage_gap	boolean		True when the utility series has a gap longer than 6 hours after the peak
customers_total	number	customers	Estimated electric customers in the county: the state's EIA-861 customer count split across counties by Census housing units
customers_reason	string		Why customers_total is empty. eia861_territory_incomplete, no_housing_units: the estimate could not be made for this county. outage_exceeds_*: the peak, after flagged readings are replaced, is above the estimate. The peak is published as reported, and the suffix says which other customer counts it also exceeds. no_outage_row: the county has no outage row, and the estimate comes with the outage run
estimate_exceeded_in_other_storm	boolean		True when this county's peak was above its customer estimate in another storm of the catalog. The share here then rests on an estimate that may be too low
customer_hours_reason	string		reporting_drop: the integral stops at a one-step collapse of the utility feed. The tail of the window is not counted
reporting_drop	boolean		True when a restore hour is empty because the utility feed fell in one 15-minute step and did not come back. The answer to give is: The utility's outage feed stopped reporting for this county partway through, so a restoration time cannot be given.

county_daily

One row per storm, county and night

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
date	date		Night of observation (the UTC date of the product file)
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
declared	boolean		True when FEMA declared the county for this storm. Same value as county_summary.declared
declared_via	string		The declaration that names the county: major_disaster, or emergency for a tropical storm's emergency declaration. A county named in both is major_disaster
in_nhc_wind_radius	boolean		True when the county lies inside the NHC 34-knot wind radius of the storm, False when it lies outside
days_from_landfall	integer	days	Days since storms.landfall (negative before). Stored nights reach back about 400 days where a prior-year baseline was read
radiance_nw_cm2_sr	number	nW/cm²/sr	Mean radiance over the valid pixels that night: the sum of radiance over valid pixels divided by their number
valid_pixels	integer	pixels	Valid pixels that night: a value, Mandatory_Quality_Flag 0, confident clear and no snow
valid_pixel_share	number		valid_pixels as a share of all pixels in the area
snow_excluded_share	number		Snow pixels as a share of pixels with a finite radiance. A different base from valid_pixel_share, so the two can add to more than 1
light_ratio	number		radiance_nw_cm2_sr divided by baseline_radiance of the summary row
thin_night	boolean		True when the night has 1 to 19 valid pixels. The radiance is kept as measured and is noisy

utility_by_county

Outage columns, one row per storm with an outage run and declared county

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
county	string		County name
state	string		State or territory postal code
declared	boolean		True: the outage run reads declared counties only
available	boolean		True when utility records cover this county's window
reason	string		Utility availability. ok: records in the window. no_rows_in_window: rows that year but none in the window. no_utility_record: no EAGLE-I rows for the county that year
source	string		utility_records when the county has EAGLE-I rows in the window, otherwise none
peak_customers_out	number		Peak customers out reported in EAGLE-I, taken after readings that hold more than one collection snapshot are replaced (see flagged_readings). A replaced reading is an interpolated value that was not measured
peak_reason	string		Why peak_customers_out is empty: no_utility_record. reason gives the cause (for example county_not_declared)
peak_customers_out_raw	number		Highest reading in the window as the EAGLE-I year file records it, before any filter
flagged_readings	integer		Readings in the window that held more than one collection snapshot and were replaced. 0 when none
peak_share_out	number		peak_customers_out divided by customers_total
peak_share_reason	string		Why peak_share_out is empty. no_utility_record: no EAGLE-I rows. no_customers_total: customers_total is empty and customers_reason says why. peak_above_customer_estimate: the peak is above the customer estimate
customer_hours_out	number	customer-hours	Customer-hours out in the window from impact_date, or up to a reporting drop when customer_hours_reason is reporting_drop
hours_to_50pct_restored	number	hours	Hours from the peak until customers out stayed at or below 50% of the peak for 6 hours
hours_to_50pct_reason	string		Why hours_to_50pct_restored is empty. not_restored_in_window: still above half the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_to_90pct_restored	number	hours	Hours from the peak until customers out stayed at or below 10% of the peak for 6 hours
hours_to_90pct_reason	string		Why hours_to_90pct_restored is empty. not_restored_in_window: still above 10% of the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_under_1pct_customers	number	hours	Hours from the peak until customers out stayed at or below 1% of customers_total for 6 hours
hours_under_1pct_reason	string		Why hours_under_1pct_customers is empty. never_above_1pct: the outage never passed 1% of customers. not_under_1pct_in_window: still above 1% at the end. no_customers_total, peak_above_customer_estimate or no_utility_record: nothing to measure against. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
coverage_gap	boolean		True when the utility series has a gap longer than 6 hours after the peak
customers_total	number	customers	Estimated electric customers in the county: the state's EIA-861 customer count split across counties by Census housing units
customers_reason	string		Why customers_total is empty. eia861_territory_incomplete, no_housing_units: the estimate could not be made for this county. outage_exceeds_*: the peak, after flagged readings are replaced, is above the estimate. The peak is published as reported, and the suffix says which other customer counts it also exceeds
estimate_exceeded_in_other_storm	boolean		True when this county's peak was above its customer estimate in another storm of the catalog. The share here then rests on an estimate that may be too low
outage_window_end_day	integer	days	Last day of the outage window, counted from impact_date: 45, or earlier at the next FEMA storm in the county
outage_window_end_reason	string		window: the full 45 days. next_storm:DR-number: the next FEMA storm in the county
window_note	string		window_cut_at_source_end: the 45-day outage window runs past 31 December 2025, the last day of the published EAGLE-I record, so records stop there
customers_source	string		EIA-861 customer table and year used for customers_total
customer_hours_reason	string		reporting_drop: the integral stops at a one-step collapse of the utility feed. The tail of the window is not counted
reporting_drop	boolean		True when a restore hour is empty because the utility feed fell in one 15-minute step and did not come back. The answer to give is: The utility's outage feed stopped reporting for this county partway through, so a restoration time cannot be given.

declared_unmatched

Declared counties with no row in the boundary file, kept with a reason

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
declared_via	string		The declaration that names the county
reason	string		merged_county_not_in_boundaries: the county was merged into another and the boundary file no longer has it. not_in_boundaries: no such county in the boundary file. outside_storm_states: in the boundary file, but its state is not in the storm's states
note	string		What happened to the county

tribal_area_summary

One row per tribal declaration storm and tribal area, the area FEMA named

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
region_scheme	string		Boundary set of the area: us-aiannh-2024, the Census 2024 American Indian, Alaska Native and Native Hawaiian areas file
region_id	string		Census area code, 4 digits. Read as text to keep the leading zero
region_name	string		Census area name
state	string		State postal code of the declaration
region_role	string		tribal_area: the row describes the tribal area FEMA named
declared_via	string		major_disaster: FEMA filed tribal declarations as major disasters
thin_region	boolean		True when the area has fewer than 20 pixels on its tiles. Its numbers are kept and are noisy
product	string		Night-light product the storm is measured with: VNP46A2 (Suomi NPP) or VJ146A2 (NOAA-20). One product per storm. Radiance from different products should not be compared between storms
baseline_radiance	number	nW/cm²/sr	Median radiance of the clear baseline nights before the storm
baseline_nights	integer		Clear nights in the baseline
baseline_source	string		event_window: baseline nights before the storm. prior_year: the same season a year earlier, used when the nights before the storm were too few. insufficient: no baseline
post_obs_nights	integer		Nights with usable pixels after the landfall date
median_valid_pixel_share	number		Median share of usable pixels after the landfall date
status	string		What the night-light series did. It is not a detected power outage. light_drop: smoothed light fell below the county's normal range soon after impact. no_measurable_outage: observed, no such fall (the name keeps an old use of the word outage for a night-light state). cloud_obscured: not observed until late and normal by then. low_signal: too noisy to judge. insufficient_data: too few nights
outage_min_ratio	number		Lowest smoothed light ratio after impact (a night-light value despite the name)
outage_min_day	integer	days	Day of that lowest ratio, counted from impact_date
days_to_50	integer	days	Days from impact until smoothed light was back at or above 50% of baseline
days_to_50_reason	string		Why days_to_50 is empty. never_below_50: a light_drop status, but light never fell below 50%. not_back_to_50_in_window: fell below 50% and did not return. no_outage_measured: observed, no light drop found (a night-light state despite the name). not_observed_cloud, not_observed_too_few_nights: the county was not observed. not_measurable_low_signal: its light is too noisy to judge
days_to_normal	integer	days	Days from impact until smoothed light was back in the county's normal range
days_to_normal_reason	string		Why days_to_normal is empty. not_back_to_normal_in_window: a light_drop status that did not return to normal. The other codes are as in days_to_50_reason
recovery_uncertain	boolean		True when a gap of 5 or more days without a clear night comes just before days_to_normal
window_end_day	integer	days	Last day of the night-light window, counted from impact_date
window_end_reason	string		window: the full window. data_end: the stored nights stop earlier. next_storm:DR-number: the next FEMA storm in the county
impact_date	date		Date the storm reached the county, day 0 of the window
impact_source	string		How impact_date was chosen. track_within_250km: first UTC date the HURDAT2 track came within 250 km of the county. track_closest_approach: the track never came that close, so the date of closest approach. catalog_landfall: storms.landfall
clear_nights_pre	integer		Clear nights before impact
clear_nights_post	integer		Clear nights after impact
method_version	string		Night-light method version
source_night_light	string		Layer marker
post_days_window	integer	days	Days after the landfall date covered by stored nights
outage_reason	string		no_county_outage_record: EAGLE-I reports counties, so there is no outage record for a tribal area
peak_customers_out	number		Peak customers out. No source reports it for a tribal area
customer_hours_out	number	customer-hours	Customer-hours out. No source reports it for a tribal area
customers_total	number	customers	Electric customers in the area. EIA-861 counts are by county
customers_reason	string		no_county_customer_count: EIA-861 customer counts are by county
peak_share_out	number		Share of customers out at the peak
peak_share_reason	string		no_county_customer_count: there is no customer count for the area

tribal_area_daily

One row per tribal declaration storm, tribal area and night

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
region_scheme	string		Boundary set of the area: us-aiannh-2024, the Census 2024 American Indian, Alaska Native and Native Hawaiian areas file
region_id	string		Census area code, 4 digits. Read as text to keep the leading zero
date	date		Night of observation (the UTC date of the product file)
declared_via	string		major_disaster, as in tribal_area_summary
days_from_landfall	integer	days	Days since storms.landfall (negative before). Stored nights reach back about 400 days where a prior-year baseline was read
radiance_nw_cm2_sr	number	nW/cm²/sr	Mean radiance over the valid pixels that night: the sum of radiance over valid pixels divided by their number
valid_pixels	integer	pixels	Valid pixels that night: a value, Mandatory_Quality_Flag 0, confident clear and no snow
valid_pixel_share	number		valid_pixels as a share of all pixels in the area
snow_excluded_share	number		Snow pixels as a share of pixels with a finite radiance. A different base from valid_pixel_share, so the two can add to more than 1
light_ratio	number		radiance_nw_cm2_sr divided by baseline_radiance of the summary row
thin_night	boolean		True when the night has 1 to 19 valid pixels. The radiance is kept as measured and is noisy

tribal_area_counties

The counties each tribal area overlaps, with the overlap as a share of each. Slivers are left out

Field	Type	Unit	Description
area_id	string		Census area code, as region_id
area_name	string		Census area name
county_geoid	string		County FIPS code, 5 digits
county_name	string		County name
county_namelsad	string		County name with its legal description (for example Coconino County)
state_fips	string		State FIPS code, 2 digits
overlap_km2	number	km²	Area of the overlap
overlap_share_of_county	number		Overlap as a share of the county's area: how much of the county lies inside the tribal area
overlap_share_of_tribal_area	number		Overlap as a share of the tribal area's area

tribal_context_county_summary

Night-light summary of each county a tribal area overlaps, for the tribal declaration storms. Context rows for the whole county

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
context_for_region_id	string		region_id of the tribal area the county row is context for
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
region_role	string		county_context: the row describes the whole county, as context for the tribal area
overlap_share_of_county	number		Overlap as a share of the county's area: how much of the county lies inside the tribal area
overlap_share_of_tribal_area	number		Overlap as a share of the tribal area's area
county	string		County name
state	string		State or territory postal code
in_nhc_wind_radius	boolean		As in county_summary
product	string		Night-light product the storm is measured with: VNP46A2 (Suomi NPP) or VJ146A2 (NOAA-20). One product per storm. Radiance from different products should not be compared between storms
baseline_radiance	number	nW/cm²/sr	Median radiance of the clear baseline nights before the storm
baseline_nights	integer		Clear nights in the baseline
baseline_source	string		event_window: baseline nights before the storm. prior_year: the same season a year earlier, used when the nights before the storm were too few. insufficient: no baseline
post_obs_nights	integer		Nights with usable pixels after the landfall date
median_valid_pixel_share	number		Median share of usable pixels after the landfall date
status	string		What the night-light series did. It is not a detected power outage. light_drop: smoothed light fell below the county's normal range soon after impact. no_measurable_outage: observed, no such fall (the name keeps an old use of the word outage for a night-light state). cloud_obscured: not observed until late and normal by then. low_signal: too noisy to judge. insufficient_data: too few nights
outage_min_ratio	number		Lowest smoothed light ratio after impact (a night-light value despite the name)
outage_min_day	integer	days	Day of that lowest ratio, counted from impact_date
days_to_50	integer	days	Days from impact until smoothed light was back at or above 50% of baseline
days_to_50_reason	string		Why days_to_50 is empty. never_below_50: a light_drop status, but light never fell below 50%. not_back_to_50_in_window: fell below 50% and did not return. no_outage_measured: observed, no light drop found (a night-light state despite the name). not_observed_cloud, not_observed_too_few_nights: the county was not observed. not_measurable_low_signal: its light is too noisy to judge
days_to_normal	integer	days	Days from impact until smoothed light was back in the county's normal range
days_to_normal_reason	string		Why days_to_normal is empty. not_back_to_normal_in_window: a light_drop status that did not return to normal. The other codes are as in days_to_50_reason
recovery_uncertain	boolean		True when a gap of 5 or more days without a clear night comes just before days_to_normal
window_end_day	integer	days	Last day of the night-light window, counted from impact_date
window_end_reason	string		window: the full window. data_end: the stored nights stop earlier. next_storm:DR-number: the next FEMA storm in the county
impact_date	date		Date the storm reached the county, day 0 of the window
impact_source	string		How impact_date was chosen. track_within_250km: first UTC date the HURDAT2 track came within 250 km of the county. track_closest_approach: the track never came that close, so the date of closest approach. catalog_landfall: storms.landfall
clear_nights_pre	integer		Clear nights before impact
clear_nights_post	integer		Clear nights after impact
method_version	string		Night-light method version
source_night_light	string		Layer marker
post_days_window	integer	days	Days after the landfall date covered by stored nights

tribal_context_county_daily

One row per tribal declaration storm, overlapping county and night. Context rows for the whole county

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
context_for_region_id	string		region_id of the tribal area the county row is context for
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
region_role	string		county_context: the row describes the whole county, as context for the tribal area
date	date		Night of observation (the UTC date of the product file)
days_from_landfall	integer	days	Days since storms.landfall (negative before). Stored nights reach back about 400 days where a prior-year baseline was read
radiance_nw_cm2_sr	number	nW/cm²/sr	Mean radiance over the valid pixels that night: the sum of radiance over valid pixels divided by their number
valid_pixels	integer	pixels	Valid pixels that night: a value, Mandatory_Quality_Flag 0, confident clear and no snow
valid_pixel_share	number		valid_pixels as a share of all pixels in the area
snow_excluded_share	number		Snow pixels as a share of pixels with a finite radiance. A different base from valid_pixel_share, so the two can add to more than 1
light_ratio	number		radiance_nw_cm2_sr divided by baseline_radiance of the summary row
thin_night	boolean		True when the night has 1 to 19 valid pixels. The radiance is kept as measured and is noisy

tribal_context_utility_by_county

Outage columns of each county a tribal area overlaps, for the tribal declaration storms. Context rows for the whole county

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
context_for_region_id	string		region_id of the tribal area the county row is context for
geoid	string		County FIPS code, 5 digits. Read as text to keep the leading zero
region_role	string		county_context: the row describes the whole county, as context for the tribal area
overlap_share_of_county	number		Overlap as a share of the county's area: how much of the county lies inside the tribal area
overlap_share_of_tribal_area	number		Overlap as a share of the tribal area's area
state	string		State or territory postal code
available	boolean		True when utility records cover this county's window
reason	string		Utility availability, as in utility_by_county
source	string		utility_records when the county has EAGLE-I rows in the window, otherwise none
peak_customers_out	number		Peak customers out reported in EAGLE-I, taken after readings that hold more than one collection snapshot are replaced (see flagged_readings). A replaced reading is an interpolated value that was not measured
peak_reason	string		Why peak_customers_out is empty: no_utility_record. reason gives the cause (for example county_not_declared)
peak_customers_out_raw	number		Highest reading in the window as the EAGLE-I year file records it, before any filter
flagged_readings	integer		Readings in the window that held more than one collection snapshot and were replaced. 0 when none
peak_share_out	number		peak_customers_out divided by customers_total
peak_share_reason	string		Why peak_share_out is empty. no_utility_record: no EAGLE-I rows. no_customers_total: customers_total is empty and customers_reason says why. peak_above_customer_estimate: the peak is above the customer estimate
customer_hours_out	number	customer-hours	Customer-hours out in the window from impact_date, or up to a reporting drop when customer_hours_reason is reporting_drop
hours_to_50pct_restored	number	hours	Hours from the peak until customers out stayed at or below 50% of the peak for 6 hours
hours_to_50pct_reason	string		Why hours_to_50pct_restored is empty. not_restored_in_window: still above half the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_to_90pct_restored	number	hours	Hours from the peak until customers out stayed at or below 10% of the peak for 6 hours
hours_to_90pct_reason	string		Why hours_to_90pct_restored is empty. not_restored_in_window: still above 10% of the peak at the end. no_customers_out: the peak is 0. no_utility_record: no EAGLE-I rows. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
hours_under_1pct_customers	number	hours	Hours from the peak until customers out stayed at or below 1% of customers_total for 6 hours
hours_under_1pct_reason	string		Why hours_under_1pct_customers is empty. never_above_1pct: the outage never passed 1% of customers. not_under_1pct_in_window: still above 1% at the end. no_customers_total, peak_above_customer_estimate or no_utility_record: nothing to measure against. reporting_drop: a one-step collapse of the feed is what crossed the threshold. peak_mismatch: the 15-minute series does not reproduce the published peak
coverage_gap	boolean		True when the utility series has a gap longer than 6 hours after the peak
customers_total	number	customers	Estimated electric customers in the county: the state's EIA-861 customer count split across counties by Census housing units
customers_reason	string		Why customers_total is empty, as in utility_by_county
estimate_exceeded_in_other_storm	boolean		True when this county's peak was above its customer estimate in another storm of the catalog. The share here then rests on an estimate that may be too low
outage_window_end_day	integer	days	Last day of the outage window, counted from impact_date: 45, or earlier at the next FEMA storm in the county
outage_window_end_reason	string		window: the full 45 days. next_storm:DR-number: the next FEMA storm in the county
customers_source	string		EIA-861 customer table and year used for customers_total
customer_hours_reason	string		reporting_drop: the integral stops at a one-step collapse of the utility feed. The tail of the window is not counted
reporting_drop	boolean		True when a restore hour is empty because the utility feed fell in one 15-minute step and did not come back. The answer to give is: The utility's outage feed stopped reporting for this county partway through, so a restoration time cannot be given.

validation

Night light against reported outages, one row per storm and night-light measure, for the storms with an outage run

Field	Type	Unit	Description
event_id	string		NHC storm identifier (for example AL092022) or FEMA disaster number (DR-4901)
measure	string		Night-light measure scored: dark_county_mean, dark_lit_ratio or dark_lit_share
n_high	integer		Declared counties with 10% or more of customers out at the peak and a value of the measure
n_low	integer		Declared counties with under 1% of customers out at the peak and a value of the measure
auc	number		Chance that a high-outage county is darker than a low-outage county on the measure. 0.5 is chance
se_hanley_mcneil	number		Standard error of auc (Hanley-McNeil)
ci_low	number		Lower end of the 95% interval for auc
ci_high	number		Upper end of the 95% interval for auc
p_unadjusted	number		Two-sided p-value of the Mann-Whitney test between high-outage and low-outage counties, so a reversed storm can count
p_holm	number		p_unadjusted after the Holm step-down correction over the scored storms of the measure
reason	string		Why auc is empty: fewer than 3 high counties, fewer than 3 low counties, or both
product	string		Night-light product the storm is measured with: VNP46A2 (Suomi NPP) or VJ146A2 (NOAA-20). One product per storm. Radiance from different products should not be compared between storms
state_group	string		The states joined to the storm's states through storms they share, space-separated. Storms with the same state_group sit in the same holdout fold

validation_pooled

Night light against reported outages pooled over storms, one row per measure and scope

Field	Type	Unit	Description
measure	string		Night-light measure scored
scope	string		Storms pooled: all, product:, kind:tropical_cyclone, kind:other, or year:
n_storms	integer		Storms with at least one scored county in the scope
n_counties	integer		Storm-county rows scored, high and low together
auc	number		Pooled chance that a high-outage county is darker than a low-outage county. 0.5 is chance
ci_hm_low	number		Lower end of the Hanley-McNeil 95% interval, counties treated as independent
ci_hm_high	number		Upper end of the Hanley-McNeil 95% interval
ci_boot_low	number		Lower end of the 95% interval from resampling whole storms. The headline uses this interval
ci_boot_high	number		Upper end of the 95% interval from resampling whole storms

validation_holdout

Pooled AUC on each fold of a grouped holdout that keeps whole groups of states together

Field	Type	Unit	Description
fold	integer		Fold number
states	string		States in the fold, space-separated
n_storms	integer		Storms in the fold
measure	string		Night-light measure scored
auc	number		Pooled AUC over the fold's storms. 0.5 is chance

Example use

import pandas as pd

# Read codes as text so the leading zeros stay.
text = {"geoid": str, "region_id": str, "context_for_region_id": str}
storms = pd.read_csv("data/storms.csv")
# Key figure: storms with night-light numbers (night_light_reason is empty).
print(storms["night_light_reason"].isna().sum())

summary = pd.read_csv("data/county_summary.csv", dtype=text)
print(summary[summary["declared"]].groupby("product")["event_id"].nunique())

# county_daily.csv is large, so read only the columns needed.
# One product per storm: compare radiance within a storm.
daily = pd.read_csv("data/county_daily.csv", dtype=text,
                    usecols=["event_id", "date", "geoid", "radiance_nw_cm2_sr"])
ian = daily[daily["event_id"] == "AL092022"]
print(ian.groupby("date")["radiance_nw_cm2_sr"].median().head())

API:

GET /v1/storms

GET /v1/storms/{event_id}

GET /v1/storms/{event_id}/counties/{fips}

GET /v1/storms/{event_id}/files

Sources and attribution

Input	Role	Terms
NASA Black Marble VNP46A2 collection 2 (VIIRS on Suomi NPP), DOI 10.5067/VIIRS/VNP46A2.002	Night light	NASA Earth science data use policy (free and open). Cite Roman, M.O. et al. (2018), NASA's Black Marble nighttime lights product suite, Remote Sensing of Environment 210, 113-143, DOI 10.1016/j.rse.2018.03.017
NASA Black Marble VJ146A2 collection 2 (VIIRS on NOAA-20), DOI 10.5067/VIIRS/VJ146A2.002	Night light, storms measured with NOAA-20	NASA Earth science data use policy (free and open). Cite Roman et al. (2018) as above
EAGLE-I county outage records, Oak Ridge National Laboratory (Figshare DOI 10.6084/m9.figshare.24237376)	Customers out, customer-hours and restore times (2014-11-01 onward)	CC BY 4.0
FEMA disaster declarations (OpenFEMA)	Which storms, counties and tribal areas are included, incident begin dates, and declared_via	Public domain
U.S. Census Bureau cartographic boundary files cb_2024_us_county_500k, cb_2020_us_county_500k and cb_2024_us_aiannh_500k (2020 and 2024 vintages)	County, Connecticut 2020 county and tribal-area boundaries. Hashes in census_boundaries_manifest.json	Public domain
NOAA National Hurricane Center best tracks: HURDAT2 Atlantic (hurdat2-1851-2025-091226.txt) and Northeast and North Central Pacific (hurdat2-nepac-1949-2025-092926.txt)	Landfall dates and times, county impact dates, 34-knot wind radii. Files in nhc_files_manifest.json	Public domain
U.S. Energy Information Administration Form EIA-861 detailed data files, https://www.eia.gov/electricity/data/eia861/	Customers by state and utility, the base of customers_total	Public domain
U.S. Census Bureau county housing unit estimates (vintages 2020 and 2024)	Splits state customer counts across counties	Public domain

How to cite

Heimdall Research (2026). Storm Blackout and Restoration (Version 1.2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23010005 Concept DOI https://doi.org/10.5281/zenodo.23010004.

@misc{storm_blackout_1_2,
  author    = {Heimdall Research},
  title     = {Storm Blackout and Restoration},
  year      = {2026},
  version   = {1.2},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23010005},
  url       = {https://doi.org/10.5281/zenodo.23010005}
}

Also cite the upstream sources listed above.

Changes

• Corrects version 1.1: the public table published the EAGLE-I file peak, including readings that hold more than one collection snapshot, and this table replaces those readings and a single-reading spike before the peak, with the file value kept in peak_customers_out_raw, Irma (AL112017) summed county peaks are 12,625,936 in version 1.1 and 8,867,374 here, Miami-Dade is 1,777,800 in version 1.1 and 898,340 here (customers_total 1,142,022.3, raw 1,777,800, flagged_readings 4), Milton (AL142024) summed county peaks are 3,803,866 and 3,736,269, and median hours to 90% restored are 37 and 52.875 for Irma and 6.75 and 58.75 for Milton. Counties listed in version 1.1 and absent from this table were undeclared (county_summary.reason is county_not_declared), or the storm has outage_reason no_utility_record_before_2014. Each of the 31 storms is in docs/CHANGES_FROM_1.1.md.
• Version 1.2 is the storm catalog: 672 FEMA-declared storms with landfall or incident dates from 2012-01-14 to 2026-08-11. Every storm has a row in storms.csv, with a reason code where it has no night-light numbers or no outage run. It follows version 1.1, the last public version.
• Night light for 658 storms: 611 by county and 47 by tribal area. 14 storms have none: 3 before_nasa_record and 11 territory_not_in_county_boundaries.
• An outage run for 477 storms, 464 of them with EAGLE-I records for at least one county. 195 storms have no outage run: 117 no_utility_record_before_2014, 30 no_published_year, 41 no_county_outage_record and 7 territory_not_in_eaglei.
• 47 tribal declarations: no county list; tribal-area rows and overlapping-county context only.
• A FEMA disaster that is the same storm as an NHC storm is not a separate row. Its number is in storms.also_known_as and the storm is counted once: AL112019 is also DR-4466 (Imelda, Texas).
• Customer-hours for counties in both version 1.1 and this table: Matthew (AL142016) sums to 140,943,783.1 in version 1.1 and 140,435,077.7 here (70 of 181 counties differ); Irma (AL112017) sums to 595,686,772.4 in version 1.1 and 586,489,025.7 here (144 of 379 counties differ).
• The 6 late-2025 storms (DR-4901, DR-4902, DR-4903, DR-4905, DR-4906, DR-4907) have an outage window that ends 31 December 2025, the last day of the published EAGLE-I record (window_note is window_cut_at_source_end).
• One product per storm (500 VNP46A2, 158 VJ146A2), named in storms.product and county_summary.product. Radiance in county_daily.csv is the mean over valid pixels.
• Declared counties of tropical storms now include the counties named in the storm's emergency declaration (declared_via is emergency).
• New tables: declared_unmatched, tribal_area_summary, tribal_area_daily, tribal_area_counties, tribal_context_county_summary, tribal_context_county_daily, tribal_context_utility_by_county.
• Columns added. storms: night_light_unit, night_light_reason, product, outage_reason, window_note, also_known_as. county_summary: declared_via, in_nhc_wind_radius, product, reporting_drop, customer_hours_reason. county_daily: declared_via, in_nhc_wind_radius, valid_pixels, thin_night. utility_by_county: outage_window_end_day, outage_window_end_reason, window_note, customers_source, reporting_drop, customer_hours_reason.
• Outage method v2. A one-step collapse of the utility feed (more than 25% of the county peak, to at most 10% of the previous reading, inside one 15-minute interval) is not a restoration. When the low spell lasts the 6-hour sustain and the feed does not return to at least half the pre-drop reading within 12 hours, the affected restore hours are empty with reason reporting_drop. A return within 12 hours is skipped and the hours are read from the later readings. Customer-hours stop at the step.
• Where the 15-minute series does not reproduce the published peak, the restore hours are empty with reason peak_mismatch and the peak is left as published.
• validation.csv has one row per storm and night-light measure, with 1,431 rows for 477 storm ids. Tables validation_pooled and validation_holdout hold the pooled scores and the scores with whole groups of states held out. County mean darkness (dark_county_mean) does not separate counties with large outages from counties with small ones: pooled AUC 0.515 over 418 storms and 5,841 counties, 95% interval 0.482 to 0.547 from resampling storms. The interval includes 0.5, which is chance. Lit-pixel radiance drop (dark_lit_ratio) separates counties with large outages from counties with small ones: pooled AUC 0.559 over 287 storms and 3,611 counties, 95% interval 0.511 to 0.605 from resampling storms. Change in the share of unlit lit-mask pixels (dark_lit_share) separates counties with large outages from counties with small ones: pooled AUC 0.566 over 287 storms and 3,611 counties, 95% interval 0.513 to 0.615 from resampling storms.

Access and contact

Access: download the files here, or use the same data through the Heimdall Research API and MCP server (heimdallresearch.com).

The tables in this package are open under CC BY 4.0. API and MCP access with a free key is free until 2026-12-31.

Corrections and questions: https://heimdallresearch.com/contact/collaborate/.