Storm Blackout and Restoration

Hurricane Milton, 15 Florida counties with the highest peak customers out (county_summary.csv)

Rows from the published Storm Blackout and Restoration 1.2 table (county_summary.csv). Estimates from public records and satellite night light. Not for real-time emergency response or for any decision about one address or property.

Estimates from public records and satellite night light. Not for real-time emergency response or for any decision about one address or property.

Milton (AL142024) on 2024-10-10. The highest peak customers out in this sample is Hillsborough County, Florida, 520308.

Milton, first 5 counties by peak customers out

Create a free account for the full result, videos and your AI assistant.

Estimates from public records and satellite night light. Not for real-time emergency response or for any decision about one address or property.

Intended use

Research on how county night light changed after U.S. storms, read next to the reported utility outages, for resilience, insurance and emergency-management studies.

Not intended for

Judging a single address or neighborhood, or real-time response. Night light is related to electricity service. It does not show when power went off or came back. See the limits below.

Counts and headline numbers use the counties FEMA declared for each storm; the files also include the other counties in the same states, marked as not declared, for comparison. 14 storms in the live list are newer than version 1.2 and not part of it. Newest storm: Champi (17 October 2026). The published version 1.2 runs through 11 August 2026. Storms after that date are not in it.

672 storms in the published table (release 1.2). 658 have night-light numbers and 477 have an outage run. 14 newer storms are in the live list and not in this table yet.

Status: what the published table holds for each storm. Night light and outage run: night-light numbers and a county outage run from utility records. Night light only: night-light numbers and no outage run. Storm row only: a row in storms.csv with a reason code and no numbers. Every row is free to download.

These are some of the columns. datapackage.json in the download lists every table and column.

Columns from county_daily.csv (one row per storm, county and night) and county_summary.csv (one row per storm and county)

Hurricane Maria, the first ten Puerto Rico municipios in alphabetical order (county_summary.csv)

Rows from the published Storm Blackout and Restoration 1.2 table (county_summary.csv). Numbers rounded to three decimals for display; the files carry full precision. none means the value is empty.

1. List every storm in the catalog

   curl https://api.heimdallresearch.com/v1/storms \
     -H "Authorization: Bearer $HEIMDALL_API_KEY"

2. One storm, with every county's summary

   curl https://api.heimdallresearch.com/v1/storms/AL152017 \
     -H "Authorization: Bearer $HEIMDALL_API_KEY"

3. One county after one storm

   curl https://api.heimdallresearch.com/v1/storms/AL152017/counties/72001 \
     -H "Authorization: Bearer $HEIMDALL_API_KEY"

   Response, trimmed

   {
     "dataset_version": "1.2",
     "doi": "10.5281/zenodo.23010005",
     "event_id": "AL152017",
     "county": {
       "geoid": "72001",
       "county": "Adjuntas",
       "state": "PR",
       "baseline_nights": 18,
       "post_obs_nights": 150,
       "status": "low_signal",
       "outage_min_ratio": null,
       "days_to_50": null,
       "days_to_normal": null
     }
   }

4. Ask an AI assistant

   Ask: “Show the county night-light reading for Adjuntas, Puerto Rico after Hurricane Maria (AL152017).” The assistant calls get_storm.

Every input, its role, and its terms

For each storm, every county in the affected states is compared night by night with its own normal light. The baseline is the median of clear, snow-free nights before the storm, leaving out nights affected by any earlier storm; if too few remain, the same weeks one year earlier are used, and the row says so. After landfall, each clear night is compared with that baseline. days_to_50 counts days from impact until smoothed light was back at or above 50 percent of baseline. days_to_normal counts days until smoothed light was back in the county's normal range. window_end_reason records why each county's window ended. Pixels flagged for cloud, poor quality, or snow are excluded, and the count of usable nights travels with every value. Each storm is measured with one satellite product, named in the product column. Radiance from different products should not be compared between storms; this release gives no conversion between them.

Outage records

Night light was compared with EAGLE-I outage records. 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.

Recomputed values

Night light for 20 storms was derived again from the night-light table, and the outage columns of 5 counties were computed again from the published EAGLE-I file. Both equal the published tables.

Known limits

Clouds and snow hide the ground, moonlight and seasonal lighting shift readings, and some light returns before power does (generators) or stays low after it returns (damaged buildings, displacement). Every value carries its observation count, and results are by county or by tribal area.

NASA publishes before-and-after light maps for single disasters. This dataset puts county night-light readings and EAGLE-I outage records for U.S. storms in one set of tables, with the method and limits published next to them.

Free

The published package is on Zenodo, Hugging Face and Kaggle, with no sign-up, and the API and MCP server with a free key.

Paid

Reports for one storm and one area. Paid products are free to every account through December 31, 2026 (midnight US Eastern). From January 1, 2027, paid plans apply only to accounts that choose one, and the free API key and free tier continue. See plans and prices.

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.

Please also cite the NASA Black Marble products.

Versions and corrections

Version 1.0 is on Kaggle and has no Zenodo DOI. Version 1.1 (31 storms with county results) is on Kaggle and has no Zenodo DOI. Version 1.2 (672 storms) has DOI https://doi.org/10.5281/zenodo.23010005 and began the Zenodo and Hugging Face copies. DOI https://doi.org/10.5281/zenodo.23010004 covers all versions. The changelog in the download records changes. Report problems through the collaboration form.

Download

Zenodo (CSV and Parquet, DOI) · Hugging Face (same files; loads with the datasets library) · Kaggle (notebook, discussion) · API and MCP with a free account. Downloads need no account.

Use notice

Estimates from public records and satellite night light. Not for real-time emergency response or for any decision about one address or property.