Cloudbursts of the Mid‑Atlantic

01 Concept map — three questions, one storm system

The paper is organized around three interlocking questions. Everything else — the radar equations, the bias corrections, the century-old bucket surveys — exists to chip away at one of these three.

How hard can it rain?The Probable Maximum Precipitation (PMP) question. Recently reframed by NASEM (2024) from "a physical ceiling" to "an extremely rare quantile of the rainfall distribution."

How hard did it rain?The measurement question. Answered here with polarimetric radar, rain gauges, and — for events before 1956 — bucket surveys and eyewitness accounts.

How does it rain hard?The mechanism question. Answered through storm dynamics (updraft/downdraft pairs, rotation, mergers) and microphysics (drop-size evolution, cold-rain vs. warm-rain processes).

02 Research question & setting

The Mid‑Atlantic region — Virginia, Maryland, and Pennsylvania — has produced some of the largest short‑duration, small‑area rainfall accumulations ever documented anywhere in the world. Two anchor records: the 1‑minute world record of 31 mm at Unionville, MD (4 July 1956), and the world record 4.5‑hour accumulation of 780 mm at Smethport, PA (18 July 1942).

The paper studies three storms from the polarimetric radar era — 30–31 July 2016 and 27 May 2018 (both Ellicott City, MD) and 8 July 2019 (Northern Virginia) — reconstructing their rainfall at 1 km² / 5‑minute resolution, then places them against a catalog of seven pre‑radar cloudbursts spanning 1819–1956, most measured by post‑storm bucket survey.

03 Data & experimental design

Instruments

WSR‑88D polarimetric radar, Sterling, VA (KLWX) — Level II volume scans, NCEI archive
Rain gauges: CoCoRaHS network (daily) + Howard/Baltimore County sub‑hourly gauges
National Lightning Detection Network — total flash density as a convective‑intensity proxy
Sterling, VA radiosonde soundings — freezing level, humidity, wind shear

Processing

K_DP computed via the Bringi method (CSU Radar Tools)
Volume scans regridded to 1 km Cartesian using Py‑ART
Rainfall fields built for a 100 km × 100 km domain per storm, centered on peak rainfall
Mean‑field bias correction: storm‑wide multiplier = Σgauge / Σradar at gauge pixels

04 The governing equations

Two power‑law rainfall estimators, switched on a threshold, plus one distribution model for the raindrops themselves.

EQ. 1 — PRIMARY ESTIMATOR (used when Z > 45 dBZ and K_DP > 0.1° km⁻¹)

R = a · K_DP^b (a = 44.0, b = 0.822)

R = rain rate (mm h⁻¹) · K_DP = specific differential phase shift (° km⁻¹). The pre‑factor a is empirically fit — Oklahoma disdrometer data optimizes to 44; Korean data optimizes to 51 (Bang et al. 2020); extreme warm rain may need ~70 (Li et al. 2023). This is the single biggest source of storm‑to‑storm bias in the paper.

EQ. 2 — FALLBACK ESTIMATOR (lighter rain / low K_DP)

R = α · Z^β (α = 0.017, β = 0.71)

Z in linear reflectivity units (mm⁶ m⁻³). The classic NEXRAD‑era Z‑R relation, used only when the K_DP conditions aren't met.

EQ. 3 — NORMALIZED GAMMA DROP‑SIZE DISTRIBUTION

N(D) = N_w f(μ) (D/D₀)^μ e^{−(3.67+μ)(D/D₀)}

N_w = normalized number concentration · D₀ = median volume diameter · μ = shape parameter. All three are inverted from (Z, Z_DR, K_DP) using the Bringi et al. (2002) method. Tracking N_w and D₀ through time is the paper's main tool for diagnosing why a storm rained hard — many small fast‑forming drops vs. fewer large drops vs. rapid alternation between the two.

THE QUIET ASSUMPTION THAT DRIVES MOST OF THE BIAS RESULTS

Both estimators implicitly assume zero vertical air motion. True rain rate is proportional to (raindrop terminal velocity + vertical air speed). In a strong wet downdraft, vertical air speed can exceed 20 m s⁻¹ — meaning a downdraft‑dominated storm can have its peak rain rate severely underestimated by both equations above, independent of any microphysics problem.

05 Interactive demonstration

Illustrative demonstration — not the paper's fitted model

Explore how the choice of K_DP pre‑factor and the presence of a downdraft each distort the rain‑rate estimate for a fixed radar‑observed K_DP. This reproduces the shape of the paper's two central error mechanisms (pre‑factor mismatch, §3–4; downdraft underestimation, §2/§4) — it is not a recomputation of any specific storm.

Observed K_DP: 3.0° km⁻¹ K_DP pre‑factor "a" — storm microphysical regime: 44 (Oklahoma / continental cold‑rain) Downdraft vertical air speed: 0 m s⁻¹ (algorithm's assumption)

—

06 Key results — the three modern storms

Peak accumulations (mm) at 1 km² by duration. Numbers in parentheses are the ratio to the NOAA Atlas‑14 1,000‑year value. Reconstructed from the paper's Table 1 for teaching purposes.

Storm	15 min	30 min	60 min	180 min	Mean bias
27 May 2018	34 (0.76)	57 (0.79)	93 (0.89)	208 (1.33)	1.08
30–31 Jul 2016	41 (0.91)	74 (1.03)	112 (1.07)	171 (1.10)	1.40
8 Jul 2019	58 (1.29)	99 (1.38)	152 (1.45)	171 (1.10)	1.60

All three storms exceed 1,000‑year values at multiple durations. The 8 July 2019 storm at McLean, VA is the single most extreme, and carries the largest correction (mean bias 1.60, with a further local/conditional correction of 2.24× at the McLean gauge itself).

Storm-by-storm mechanism

27 May 2018 — Ellicott City

Cold‑rain signature: hail aloft (Z > 68 dBZ at 5 km), large Z_DR updraft column
Two cells merged 2012–2017 UTC → strong updraft → strong downdraft (descending K_DP core)
Storm rotation peaked 1.5 m s⁻¹ km⁻¹ at 2021 UTC, coincident with the merger
Conditional bias worst during the updraft→downdraft transition — consistent with the vertical‑motion mechanism

30–31 Jul 2016 & 8 Jul 2019

More warm‑rain dominated: high drop number density, smaller drops, collision‑coalescence growth
Standard pre‑factor of 44 (tuned for continental cold rain) underestimates these storms
July 2019: rain rates at McLean reached 198–280 mm h⁻¹ over a 30‑min window
Dead Run flood: runoff ratio ≈ 1.0 (essentially all rain became streamflow)

07 The pre‑radar catalog — and Smethport, PA

Seven historical cloudbursts, 1819–1956, nearly all measured by post‑storm bucket survey rather than instrument:

Date & place	Accumulation	Duration
26 Jul 1819, Catskill, NY	~305 mm	30 min
5 Aug 1843, Concord, PA	~406 mm	3 hr
19 Jul 1889, Rockport, WV	~483 mm	2 hr 10 min
24 Aug 1906, Guinea, VA	235 mm	40 min
4–5 Aug 1943, Little Kanawha, WV	>380 mm	2 hr
4 Jul 1956, Unionville, MD	31 mm	1 min (rain gauge, world record)
18 Jul 1942, Smethport, PA	780 mm (world record)	4.5 hr

Smethport also produced 165 mm in 10 minutes (≈1,000 mm h⁻¹) and 240 mm in 25 minutes (≈580 mm h⁻¹), plus 400+ individual bucket‑survey points and the regional envelope‑curve flood peak — still unmatched by any modern radar‑era storm in this dataset.

Cross-cutting physical threads the authors identify

Extreme convective intensity in every account — severe/continuous lightning and thunder, in both 1819 and 2019 alike
Storm rotation — tornado at Concord (1843), 18 hail reports at Smethport (1942)
"Streams and sheets" language — repeated eyewitness description of rain not falling as drops, echoing the modern DSD evolution (large→small drop transitions) documented via radar
Storm mergers — described at Rockport, Catskill, Concord, and Smethport, mirroring the 27 May 2018 cell‑merger mechanism

08 Limitations & failure modes

Conditional bias, not just mean bias. A single storm‑wide multiplier does not fix underestimation concentrated at the peak — local, gauge‑cluster corrections shifted the May 2018 3‑hr peak from 208→240 mm and the July 2019 McLean 30‑min value required a 2.24× local correction on top of the 1.60× storm‑wide one.
Pre‑factor is not universal. The value of 44 in R=aK_DP^b is an Oklahoma‑optimized constant; it systematically underestimates warm‑rain‑dominated storms.
Zero vertical‑motion assumption. Neither rainfall estimator accounts for downdraft‑enhanced fall speed; this is treated as a first‑order (not secondary) error source.
Sparse ground truth at the true extreme. The July 2019 McLean 30‑min total is 1.7× the historical maximum ever recorded in Maryland's 1890s‑to‑present gauge network — the climatological network is too spatially sparse to ever sample the true extreme.
Bucket surveys can't be re‑QC'd. The entire pre‑1956 record (except Unionville) rests on post‑hoc container surveys and eyewitness corroboration — invaluable, but not instrument‑traceable in the modern sense.

09 Source‑supported conclusions vs. project inference

Source‑supported (the authors' claims)

All three modern storms exceed 1,000‑yr return values at multiple durations, yet remain modest next to historical cloudburst maxima
Conditional bias is a prominent, physically explainable feature — not random noise — tied to microphysical regime and vertical motion
PMP is better framed as an extreme quantile than a physical ceiling (per NASEM 2024); physical process understanding is the necessary path to characterizing that tail
The Smethport, PA 1942 storm warrants dedicated modern reanalysis given its unmatched observational density

Project inference (not claimed by the authors)

Smethport's 165 mm/10 min and 780 mm/4.5 hr totals are fixed points any PA historical extreme‑precipitation reconstruction should remain consistent with, even pre‑dating gridded products like PRISM
Conditional bias scaling with storm microphysical regime implies gridded/reanalysis QPE bias is unlikely to be uniform across storm types — a mechanism worth testing against any PA/Mid‑Atlantic QPE discontinuity investigation
The paper's quantified treatment of vertical‑motion‑driven underestimation is a concrete, underused mechanism worth flagging in a standing search for bold/underused physical‑statistical tools in precipitation science

A falsification test the paper gestures at but doesn't run

If the K_DP pre‑factor really is storm‑type dependent (≈44 for continental cold rain, 50–70 for warm‑rain‑dominated storms), then classifying storms by microphysical regime before selecting a pre‑factor — rather than applying one fixed value everywhere — should measurably shrink both mean‑field and conditional bias. This is directly testable against any disdrometer‑equipped radar domain with a hail‑report archive to sort storms by.

10 Glossary

PMP (Probable Maximum Precipitation): Rainfall depth used to design high‑hazard dam spillways; recently redefined by NASEM (2024) as an extremely low‑probability quantile rather than a physical ceiling.
K_DP (specific differential phase shift): Range derivative of the phase difference between horizontally and vertically polarized radar returns; strongly tied to liquid water content, comparatively insensitive to hail contamination.
Z_DR (differential reflectivity): Polarimetric variable sensitive to raindrop size/shape; a "Z_DR column" extending upward marks a strong updraft carrying large drops aloft.
Mean‑field bias: A single multiplicative correction applied storm‑wide: Σ(gauge totals) / Σ(co‑located radar totals).
Conditional bias: Bias that varies systematically with rainfall intensity itself — typically worst exactly at the peak, which a flat mean‑field correction cannot fix.
Normalized gamma DSD: Statistical model of the raindrop size distribution, parameterized by N_w (number concentration), D₀ (median volume diameter), and μ (shape).
Storm‑relative area correction factor: Ratio of maximum rainfall at a larger area (e.g. 100 km²) to the 1 km² peak — a measure of how "peaky" vs. "broad" a storm's rain field is.
Runoff ratio: Total streamflow volume divided by total rainfall volume for a watershed and storm; near 1.0 means almost all rain immediately became runoff.

11 Full narrative — listen or read

The complete ~4,700‑word / ~30‑minute TTS narrative, embedded below with browser playback controls. Six chapters, matching the sections above.

Not playing.

Show full transcript