harshithkethavath/FloodPrint-FFW-Labeled-WS2024

FloodPrint: NWS Flash Flood Warning Verification Labels (Warm Season 2024)

Part of the FloodPrint Collection — a CONUS-scale dataset for NWS Flash Flood Warning verification research.

Dataset Summary

This is the primary research-ready output of the FloodPrint pipeline. Each record is a NWS Flash Flood Warning polygon annotated with a binary verification label — whether the warning was confirmed by at least one Flash Flood Local Storm Report (LSR) falling within the warning polygon during its valid time window.

This dataset is the Phase 3 output and the direct input to any downstream classification or spatial analysis. It is derived from a spatiotemporal join of FloodPrint-LSR-WS2024 and FloodPrint-FFW-Polygons-WS2024.

Headline finding: the national verification rate for CONUS warm season 2024 is 45.2% — substantially lower than the 70–85% rates commonly reported in the literature, with a strong geographic gradient driven by climate regime and spotter network density.

Dataset Details

Property	Value
Record count	~3,700 (one per warning polygon)
Geographic coverage	CONUS
Temporal coverage	April 1 – September 30, 2024
Verification rate (national)	45.2%
Spatial representation	Polygon geometries (Shapely, WGS84 / EPSG:4326)
Source	Spatiotemporal join of IEM LSRs and IEM WWA polygons
File format	Parquet (geometry stored as WKB)

Data Fields

Column	Type	Description
eventid	str	Unique NWS event identifier
wfo	str	3-letter NWS Weather Forecast Office that issued the warning
issue_utc	datetime64[UTC]	Warning issuance time (UTC)
expire_utc	datetime64[UTC]	Warning expiration time (UTC)
geometry	geometry (Polygon)	Warning polygon boundary in WGS84 (EPSG:4326)
verified	bool	True if at least one LSR fell within the polygon during the valid window; False otherwise

Reading geometry from Parquet

import pandas as pd
import geopandas as gpd
from shapely import wkb

df = pd.read_parquet("ffw_labeled_ws2024.parquet")
df["geometry"] = df["geometry"].apply(wkb.loads)
gdf = gpd.GeoDataFrame(df, geometry="geometry", crs="EPSG:4326")

# Verification rate
print(gdf["verified"].mean())  # 0.452

Verification Methodology

The spatiotemporal join follows two criteria, both of which must be satisfied for a warning to be labeled verified=True:

Spatial criterion

An LSR is considered spatially associated with a warning if a 5 km buffer around the LSR point intersects the warning polygon boundary (not just the interior). This accounts for the spatial uncertainty inherent in LSR location reporting, particularly in rural areas.

Temporal criterion

An LSR is considered temporally associated with a warning if its valid_utc falls within the window [issue_utc, expire_utc + 60 minutes]. The 60-minute grace period post-expiration is a deliberate design choice: flash flood events routinely continue to develop and generate LSRs after formal warning expiration, and excluding these would systematically undercount true verification.

A warning is labeled verified=True if at least one LSR satisfies both criteria simultaneously.

Key Findings

Finding	Value
National verification rate	45.2%
Desert Southwest WFOs (e.g. PSR, TWC, EPZ)	0–13%
Northeast WFOs (e.g. ALY, BOX, PHI)	77–85%
Primary driver of geographic gradient	Climate regime + LSR spotter network density

The 45.2% national rate is substantially lower than the 70–85% rates commonly cited in the NWS verification literature. The dominant explanation is spotter density bias: arid and sparsely populated regions have fewer spotters, so verified flash flood events go unreported as LSRs even when they genuinely occur. The measured verification rate in these regions reflects observation network coverage as much as true meteorological warning skill.

Known Limitations

• LSR spotter bias: Verification rates in rural and arid regions are likely underestimates of true meteorological skill. The label verified=False does not necessarily mean no flooding occurred — it means no LSR was filed.
• One warm season: Results reflect 2024 conditions only. Inter-annual variability (ENSO state, precipitation anomalies) may shift regional verification rates.
• Original polygon only: Amended or cancelled warnings use the original issued polygon geometry (standard NWS verification methodology).
• Binary label only: This dataset provides a binary verification label. Continuous targets (false alarm area fraction, lead time to first LSR) are not included in this release.

Citation

If you use this dataset, please cite the FloodPrint project and the IEM data source:

@dataset{floodprint_ffw_labeled_ws2024,
  author    = {Kethavath, Harshith},
  title     = {FloodPrint: NWS Flash Flood Warning Verification Labels (Warm Season 2024)},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/harshithkethavath/FloodPrint-FFW-Labeled-WS2024}
}

Original data source: Iowa Environmental Mesonet (IEM), Iowa State University. https://mesonet.agron.iastate.edu

License

Creative Commons Attribution 4.0 International (CC BY 4.0) — source data is publicly available from NOAA/IEM with no use restrictions.