Severe thunderstorm(+) Climatology and Trends

Hazard

Threshold

Place

County

From

Smoothing

Loading…

Climatology

drag · scroll to zoom · click a state to center, a county to pick it

Trend

Trends are inherently far noisier: focus on coherent spatial patterns, not the value in any single county.

both maps share one view

Methods

Sources. Hail, tornado and thunderstorm wind come from the NOAA/NCEI Storm Events Database, using county-type records (CZ_TYPE = C). Freezing rain comes from NOAA/NCEI’s GHCNh (Global Historical Climatology Network-hourly, the successor to the Integrated Surface Database): the hourly and special reports (METARs) of airport weather stations across the lower 48. Peak wind comes from the Integrated Surface Database for Indiana stations. The remark FZRANO (freezing-rain sensor not operational) is not counted as freezing rain, and a peak-wind group with a three-digit speed (100 kt or more) is treated as a coding error.
Extreme heat comes from the ECMWF ERA5 reanalysis (hourly, 0.25°, NCAR’s copy), not from reports. Two hazards: Extreme Heat (temperature) counts days whose maximum 2 m air temperature reached 90, 95, 100 or 105 °F; Extreme Heat Index counts days whose maximum heat index reached 90, 100, 105, 110 or 115 °F, with heat index computed every hour from temperature and dewpoint by the NWS formula (the Rothfusz regression and its humidity adjustments). A day is midnight to midnight local standard time. Each ERA5 land cell counts its own days, and a county’s value is the area-weighted mean of the land cells it overlaps (ERA5 land–sea mask ≥ 0.5; sea and lake cells carry marine temperatures and would pull coastal counties down; the few island counties with no land cell take the nearest one). State and national panels are the area-weighted mean over counties, i.e. the days a typical place saw (so values need not be whole numbers). The method reproduces the Indiana heat-index analysis it grew out of to within 2 point-days in 86,000, and an independent recomputation from the raw hourly data matched it. Lower 48 only; 1950 onward. The newest month’s last day lacks a few overnight hours until the following month is out, and is recomputed then. Before about 1960 the top heat-index thresholds (110 and 115 °F) are unreliable: early ERA5 has occasional one-hour humidity spikes (an 88 °F dewpoint in Nebraska, say) that the heat index turns into extreme values, so those early counts run high and trends there are understated. Contains modified Copernicus Climate Change Service information; ERA5 is produced by ECMWF.
Heavy rain comes from PRISM daily precipitation (4 km grid, PRISM Group, Oregon State University), obtained through ACIS, which counts the days on its server so no daily grids need downloading. A day is the 24 hours ending 12 UTC; totals are taken to the nearest hundredth of an inch, as rain is reported, and a day counts at 1, 2 or 3 inches. Each PRISM cell counts its own days; a county’s value is the area-weighted mean of the cells it overlaps, and state and national panels the area-weighted mean over counties — the days a typical place saw. 1981 onward (PRISM’s daily record), lower 48 only. Heavy rain opens on 2002 because PRISM itself changes then: from 1 January 2002 its daily precipitation uses weather radar (NWS Stage II/IV, then MRMS from July 2020) to place rain between the gauges across the central and eastern U.S., and PRISM warns that this adds non-climatic variation to trends. Against long-record gauges, heavy-rain days jump at 2002 in every eastern state we checked (≥ 2″ days go from about 0.6 to 0.86 of the gauge count; no jump in the West, where radar is not used), so a trend that spans 2002 is inflated. Earlier years can still be chosen in the year boxes. The last six months are provisional and are fetched again each month as PRISM revises them. PRISM is built from the gauges that reported each day, and that network changes over time, so trends carry some uncertainty from it. The method reproduces the Indiana heavy-rain analysis it grew out of exactly, cell by cell.
Heavy snow comes from daily snowfall at NOAA Cooperative Observer and first-order stations (GHCN-Daily, via ACIS) — measured by a person with a ruler, since no gridded product estimates daily snowfall well enough for trends. A day counts at 1, 4, 6 or 12 inches of new snow; seasons run July–June, labeled by their ending year, 1950–51 onward. Quality control follows Kunkel et al. (2009), automated: blank snowfall on days that could not have had snow (no precipitation, or a minimum above 32 °F) is set to zero; days with more than 60 in, or with snow but exactly no precipitation, are set to missing; a season counts only if October–May is at least 90% complete and the observer actually reported snowfall at least once in each of December, January and February (so a station that had stopped measuring snow gets no invented zero seasons), and is dropped as estimated when most of its snowfalls read exactly 10:1 from the gauge (the EPA snowfall indicator’s screen). A station appears with at least 30 valid seasons. The default QC stations set further requires at least 60 of 76 seasons, five in each of the first and last decades, and no break or drift against its neighbours (a Pettitt test and a trend test on the ratio to a neighbour reference, in place of Kunkel’s panel of experts). CoCoRaHS is left out: its records mostly start after 2000. The lower panels average the stations observed in at least 90% of the seasons in the window, as departures from each station’s own mean (added back to their average), so a station missing a season does not shift the series.
Locate each report (storm hazards). Storm Events reports are assigned to a county by STATE_FIPS + CZ_FIPS rather than by county name, which avoids name collisions and independent cities. Connecticut is drawn with its eight historical counties, because Storm Events still files reports under them rather than the planning regions the Census adopted in 2022. Puerto Rico reports are included from 2009, the first year they carry municipio codes (earlier ones use forecast zones). Station hazards stay at their station.
Apply the threshold. Each report is scored against the selected size or intensity cut (for extreme heat, each day’s maximum against the temperature or heat index cut). Thresholds are cumulative — a 2″ hailstone also counts as a 1″ one.
Collapse to hazard days (storm hazards). Count distinct calendar dates on which a county — or a station — saw at least one qualifying report. One storm producing forty reports in a county on one afternoon is one hazard day.
Screen the station records. A station-year below 90% complete is treated as no observation and dropped, not as zero days. That is the opposite of the county rule below, and both are right: every county exists every year, a station does not. For freezing rain, QC stations applies the criterion of DelPizzo et al. (2025), who kept only sites with an overnight human observer. Their list covers the Great Lakes; nationally the page applies the same test from the record itself: a site qualifies if, before automation (1985, 1987 or 1990), it reported in the small hours (1–5 am local) on at least 90% of its days. That reproduces their list — 75 of their 76 U.S. sites pass, and the 76th, Champaign, is kept as theirs — and gives about 550 of the roughly 950 stations nationally (every federal ASOS, plus military fields with their own observers). The test describes 1985–90, not today: many of these sites are now automated too. Freezing rain opens on that subset. In this record the screen matters less than expected: unscreened stations average about 5% more freezing-rain days than a screened station nearby (median of 165 pairs within 60 km), though more in 2001–05 (11–26%) than since, a sign that automated freezing-rain reporting changed in the early 2000s. A day is a UTC day (7 pm to 7 pm EST), as in the METAR record. Seven mountain and pass sites where most “FZRA” hours came with freezing fog or visibility at or below half a mile — ice from cloud rather than rain (Mount Washington, Sexton Summit, Stampede Pass, Blue Canyon, Mullan Pass, Guadalupe Mountains, Bryce Canyon) — are left out. Peak wind is measured differently and is not screened this way, so it opens on all stations.
Store those counts per county (or station) × year × month × threshold. Everything below is computed live in the browser from that table.
Climatology map. Mean annual hazard days over the selected period. For counties a year with no days is a real zero and is counted in the denominator; for stations, unobserved years are excluded rather than counted as zero.
Trend map. Ordinary least squares on the annual counts, reported per decade. A county is left gray (a station hollow) unless it clears the same three coverage gates as the 2° grid page: enough years carrying a day (at least 8, or 30% of the period), those years spanning at least 70% of the period, and at least 2 of them in each half — for a station counted over its observed seasons, which must also cover at least half the period. The annual and monthly panels apply the same gates. Significance is a two-sided t test on the slope; stations reaching p < 0.05 are ringed.
Annual and monthly panels. For a selected county, that county alone. For a state, calendar dates on which anywhere in the state qualified — deliberately not the sum of county-days, since one storm can cross twenty counties in an afternoon. For extreme heat, a state (or the country) is instead the area-weighted mean over its counties: the days a typical place there saw. For stations with none selected, the mean over those observed in every year of the period, so a station entering or leaving the record cannot masquerade as a trend.
Smoothing. The two county maps are drawn through a light Gaussian blur, sigma 2°, on by default and switchable to Raw. Measured over the real county centroids that means half the weight comes from within 230 km and 90% from within 381 km, with a hard cutoff at 444 km, so a point smears into a blob roughly 520 km across and a median of 286 counties contribute to each value. That is a firm blur: it resolves the broad corridors rather than county-scale structure, and the peak hail value drops from about 15 to 6 days/yr. Switch to Raw to see the underlying county numbers. The 2° grid page uses the same sigma. It is area-weighted — without that, Georgia's 159 small counties would outvote a single large western one covering the same ground — and normalized, so only counties that have a value contribute and coastal counties are not dragged towards zero by the empty space beyond them. The smoothed trend map is the trend of the smoothed field: every county's own slope is blurred, gated or not (an ordinary least-squares slope is linear in the data, so this equals the slope of the smoothed annual series), and the trend gates are then applied to that smoothed series. A county too sparse for a trend of its own therefore shows its neighbourhood's, and goes gray only when the whole neighbourhood is too sparse; hover gives the county's own trend, or says it has none. It never touches the annual or monthly panels. Because it is a neighborhood average, the drawn value is not the county's own — hover reports both when they differ. Station hazards and derecho events are not smoothed: those are points and single storms, not fields. Extreme heat is never smoothed either: ERA5 is already a continuous gridded field. Heavy rain is smoothed at 1° rather than 2° by default, switchable to Raw: heavy rain is patchy over much shorter distances than heat, so single-county trends in particular are noisy, while PRISM resolves real structure at 4 km that the 2° blur would wash out. For heavy rain only the trend map is smoothed; its climatology is always drawn raw, because a long-term mean is not noisy and a blur moved western rain shadows badly (Grant County, WA: 0.02 → 2.9 days/yr at ≥ 1″).
Color scales. Set nationally (all counties, or every station shown), whatever part of the map is on screen: robust upper percentiles for climatology (the 98th for the snowfall and freezing-rain station networks, whose few mountain sites would otherwise bleach every other dot), symmetric about zero for trend. When values run past the top of the climatology scale its top label reads ≥; hover a county for its own number.
Derecho is the same county climatology and trend as thunderstorm wind, but counting only wind reports that belong to a derecho. A county scores a derecho event when it has a qualifying wind report inside the coherent wind swath of a definitive derecho in the SPC archive — NEXRAD-confirmed cases only, so the record begins in 1996. A county the swath crossed without a report is not credited (about a third of the counties at least half inside the swath polygons), so the map understates swath coverage, and early swaths whose reports often carry no speed (1996–2002) lose the most. Derechos here are warm-season storms: all 96 definitive cases fall in April–August. It is counted in events per year rather than hazard days, because a derecho is one coherent storm rather than a category of report. Each event is dated by the swath's own start (UTC) rather than each report's local date: half these swaths cross a date boundary, and dating by report counted one storm twice in any county hit either side of midnight. Two swaths that start on the same date count once in a county both cross, so nationally the 96 definitive swaths fall on 93 dates. A report belongs to a swath when it falls inside the archive's UTC window, keeps pace along the track with the clock, lies no farther from either endpoint than the swath's length plus 100 km (the paper's three tests), and — our addition — lies within 300 km of the straight start-to-end track line, which keeps separate storms far off to the side out of the swath. Values are small by construction — the busiest counties see about one derecho every two years — so read the map as a corridor map, not a frequency map, and treat trends over a record this short with the caution the yellow note above the trend panel describes.
Derecho events — individual swaths, their tracks and reports — are on the historical events page, with the archive’s confidence tiers.