sachinkg12/us-county-hazard-features

US County Multi-Hazard Features Dataset

A curated, ML-ready dataset of 1,014,930 county-month observations spanning 3,222 US counties from 2000 to 2026, integrating 7 federal data sources into 42 engineered features for predicting FEMA disaster declarations 90 days in advance.

Dataset Summary

Property	Value
Rows	1,014,930
Columns	50 (42 features + target + metadata)
Counties	3,222 (all US counties with available data)
Time span	2000-01 to 2026-03
Granularity	County-month
Target	declaration_next_90d — binary, FEMA disaster declaration within 90 days
Positive rate	11.02%
Format	Apache Parquet (Snappy compression)

Data Sources

This dataset integrates 7 federal data sources — all free, no API keys required:

Source	What
FEMA Disaster Declarations	Every federal disaster declaration by county (1953–present)
USGS Earthquakes	M2.5+ seismic events with coordinates (1964–present)
NOAA Storm Events	Tornadoes, floods, hurricanes, hail + casualties/damage (2000–present)
US Census Bureau	County demographics, housing, economics
US Drought Monitor	Weekly drought severity by county, D0–D4 (2000–present)
NIFC Wildfires	Wildfire incidents with acres burned (2000–present)
NFIP Flood Claims	National Flood Insurance Program claims and payouts (1978–present)

Features (42 total)

FEMA History (7)

Rolling window declaration counts and recency metrics.

Feature	Type	Description
declarations_1yr	int	Disaster declarations in prior 1 year
declarations_3yr	int	Disaster declarations in prior 3 years
declarations_5yr	int	Disaster declarations in prior 5 years
declarations_10yr	int	Disaster declarations in prior 10 years
months_since_last_decl	int	Months since most recent declaration (-1 if none)
major_disaster_ratio	float	Fraction of declarations that were major disasters
ia_program_ratio	float	Fraction with Individual Assistance programs

Storm Events (10)

NOAA severe weather aggregations.

Feature	Type	Description
storm_event_count_1yr	int	Storm events in prior 1 year
storm_event_count_5yr	int	Storm events in prior 5 years
storm_deaths_5yr	int	Storm-related deaths in prior 5 years
storm_injuries_5yr	int	Storm-related injuries in prior 5 years
storm_property_damage_5yr	float	Property damage ($) in prior 5 years
storm_crop_damage_5yr	float	Crop damage ($) in prior 5 years
tornado_count_5yr	int	Tornado events in prior 5 years
flood_count_5yr	int	Flood events in prior 5 years
hail_count_5yr	int	Hail events in prior 5 years
max_tor_f_scale_5yr	int	Maximum tornado F-scale in prior 5 years

Socioeconomic (5)

US Census demographic and economic indicators.

Feature	Type	Description
population	long	County population
housing_units	long	Number of housing units
median_home_value	long	Median home value ($)
population_density	float	People per square mile
land_area_sq_mi	float	County land area in square miles

Drought (4)

US Drought Monitor severity metrics.

Feature	Type	Description
drought_severity_avg_5yr	float	Average drought severity score (5yr)
drought_max_severity_5yr	float	Maximum drought severity score (5yr)
severe_drought_weeks_5yr	int	Weeks of severe drought (D2+) in 5 years
drought_d4_pct_max_5yr	float	Peak percentage of county in D4 (exceptional) drought

Wildfire (4)

NIFC wildfire incident metrics.

Feature	Type	Description
wildfire_count_1yr	int	Wildfire incidents in prior 1 year
wildfire_count_5yr	int	Wildfire incidents in prior 5 years
wildfire_acres_burned_5yr	float	Total acres burned in prior 5 years
wildfire_max_acres_5yr	float	Largest single wildfire (acres) in 5 years

NFIP Flood Insurance (3)

National Flood Insurance Program claim patterns.

Feature	Type	Description
nfip_claim_count_5yr	int	NFIP claims in prior 5 years
nfip_total_payout_5yr	float	Total NFIP payouts ($) in prior 5 years
nfip_avg_payout_5yr	float	Average NFIP payout ($) in prior 5 years

Spatial (2)

Neighborhood and state-level context.

Feature	Type	Description
neighbor_avg_declarations_5yr	float	Average 5yr declarations of neighboring counties
state_avg_declarations_5yr	float	Average 5yr declarations across the state

Cascade Interaction Features (7)

Novel contribution: Multiplicative interaction terms capturing multi-hazard co-occurrence.

Feature	Type	Description
cascade_drought_fire_risk	float	drought_severity_6mo × log1p(wildfire_acres_1yr)
cascade_fire_flood_risk	float	log1p(burn_scar_acres_18mo) × flood_events_1yr
cascade_hurricane_flood_risk	float	hurricane_declarations_60d × flood_events_30d
cascade_earthquake_landslide_risk	float	significant_quakes_90d × severe_storms_30d
cascade_storm_compound_count	int	Severe storms in prior 30 days (compound events)
cascade_active_chains	int	Count of active cascade interactions (0–5)
cascade_max_chain_length	int	Longest active hazard chain (1–3)

Target & Metadata

Column	Type	Description
fips	string	5-digit FIPS county code
year_month	string	Observation month (YYYY-MM)
declaration_next_90d	bool	Target: FEMA declaration within 90 days
declaration_type_next_90d	string	Declaration type if positive (DR, EM, etc.)

Usage

import pandas as pd

df = pd.read_parquet("us-county-hazard-features.parquet")

# Temporal train/test split (recommended)
train = df[df["year_month"] < "2022-01"]
val = df[(df["year_month"] >= "2022-01") & (df["year_month"] < "2023-01")]
test = df[(df["year_month"] >= "2023-01") & (df["year_month"] <= "2024-12")]
# Note: Exclude months after 2024-12 — FEMA declaration data is incomplete

# Feature columns (42 features, no temporal — see paper for ablation justification)
FEATURE_COLS = [c for c in df.columns if c not in [
    "fips", "year_month", "declaration_next_90d", "declaration_type_next_90d",
    "month_of_year", "is_hurricane_season", "is_tornado_season", "is_wildfire_season"
]]

Benchmark Results

Using XGBoost with temporal split:

Model	ROC-AUC	PR-AUC	F1
Naive (prior)	0.500	0.549	0.000
Logistic Regression	0.542	0.139	0.199
Random Forest	0.845	0.287	0.334
XGBoost	0.893	0.555	0.482

95% Bootstrap CI: ROC-AUC [0.890, 0.896]

Key Findings

1. FEMA Dominance: Removing FEMA features drops AUC from 0.89 to 0.63 — declaration history is the strongest predictor, suggesting the federal process is path-dependent.

2. Cascade Interactions: Multi-hazard cascade features improve compound disaster detection (recall lift +2.9% for cascade events, ROC-AUC 0.907 vs 0.893 overall).

3. Declaration Equity: Low-income counties (Q1) show 2.3x higher prediction residuals than wealthy counties (Q4) at the same hazard exposure level (p < 1e-100), suggesting structural inequities in federal disaster declarations.

Citation

@dataset{gupta2026uscountyhazard,
  title={US County Multi-Hazard Features for Disaster Declaration Prediction},
  author={Gupta, Sachin},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/sachinkg12/us-county-hazard-features}
}

License

Apache 2.0

Source Code

HazardCast — Full pipeline: data ingestion, feature engineering, model training, and REST API.