Crop Damage: Nature's Number-One Offender

Dek: Since 1993 the National Weather Service has recorded $47 billion of crop damage across 30,825 storm events. Hail — the peril the crop-insurance and storm-restoration trades are built around — accounts for 9% of it. Drought and freeze together account for 79%. Crops are imperiled by a relatively slow-moving force.

---

Abstract

Severe-weather damage to crops is usually pictured as something that arrives in an afternoon: hail flattens a field, wind lodges a stand of corn, a tornado cuts a swath. Counted across the full measured record, that picture is close to inverted. Using the TerraPulse peril Eventdexes — 1,806,211 recorded US storm events drawn from NOAA's National Centers for Environmental Information Storm Events Database — we rank the perils by the crop damage recorded against them. The event record reaches back to 1950, but no crop damage is recorded before 1993: of 184,673 events predating that year, not one carries a figure above zero. The damage record is therefore 1993–2026, and for ten of the thirteen perils it begins in 1996. Within it, 30,825 events carry a crop-damage figure above zero, totalling $46,951,101,773.

Drought accounts for 62.0% of that total and freeze events a further 16.8%; hail is third at 9.3% and wind fourth at 6.1%. The four together are 94.2% of all recorded crop loss. The gap is not in how often the perils strike but in what each one costs when it does. Because NCEI files one row per zone or county, the independent unit is the storm episode, not the row: across 550 damaging drought episodes and 3,579 hail episodes, the median drought episode records $3,127,500 of crop loss against $45,000 for hail — a factor of 69.5, with Cohen's d on log₁₀ values of 1.38. We test the null hypothesis that the per-event loss distribution is the same for every peril and reject it (Kruskal-Wallis H = 5,542, df = 12, p ≈ 0, ε² = 0.18).

Two things this is not. It is not a claim about physical severity: NCEI damage figures are estimates entered by National Weather Service offices at the time of reporting, not audited insurance claims. And it is not a like-for-like per-event comparison: drought and freeze are recorded against forecast zones over months, hail against counties over minutes. Both limits are quantified below rather than waved at, and the ranking survives both.

---

1. Why ask this

The industries that exist to price and repair weather damage to crops are organised around sudden perils. Hail is the peril that has a claims-adjustment trade attached to it, an aerial-imagery industry, a per-address verification market. Ask which weather event costs American agriculture the most and hail is the intuitive answer.

The measured record has been available to check that for decades. NCEI's Storm Events Database records, for each event its offices logged, a crop-damage figure alongside the property-damage figure. We are not aware of a published per-peril ranking of that field, though we make no claim to have surveyed the literature exhaustively; what follows is simply the sum, by peril, of what the offices recorded.

Prior TerraPulse work has used this drought record twice, and neither study asked this question. Drought–Wildfire–AQI Cascade (issue #48) tested whether drought propagates into wildfire and then air quality, and returned a null: no detectable drought–AQI link in routine data, with the apparent signal traced to a seasonal confound. ENSO Phase vs US Drought Severity remains a draft pending a deeper backfill and reports no result. Both treated drought as a driver of something else. This paper treats it as a cost, which is a different question of the same dex.

So we did. The question is deliberately narrow and entirely descriptive: across everything the National Weather Service has recorded, which perils carry the crop-damage dollars, and how much does each one cost per event?

2. Data

The peril Eventdexes. Each slot is one recorded storm event, filed by phenomenon. We use fourteen kinds — drought_report, cold, hail, wind, heavy_rain, wildfire, heat, winter, lightning_report, dust, coastal, fog, avalanche, and tor — plus flood_report. Every one traces to NCEI Storm Events; tor additionally cross-matches the SPC tornado history.

1,806,211 events; damage recorded 1993–2026. Of these, 1,137,853 (63.0%) carry a crop-damage value at all, and 30,825 carry one above zero — every one of them since 1993.

Coverage is quoted for 1996 onward. Raw coverage is not comparable across perils: hail, wind and tornado carry event records stretching back to 1955 from an era that logged crop damage as a flat zero, which lifts their apparent coverage against perils whose record starts in 1996. Restricted to 1996+, hail's coverage is 50.8% rather than the 60.3% a whole-record count gives, wind's 60.2%, tornado's 57.6%. Drought (53.8%) and freeze (57.3%) have no pre-1996 rows and are unaffected.

The independent unit is the episode, not the row. NCEI files one row per zone or county affected. Damaging drought rows cluster 8.5 to an episode, freeze rows 6.5, hail rows 3.2. Row counts are therefore reported for description; every significance test below uses episodes.

Measured events, estimated losses. This distinction is the paper's most important caveat and we state it before any result. The events are measured — an office recorded that hail fell here, at this time, of this size. The damage figures are estimates made at reporting time, not audited claims settled later. They are the best public per-event record that exists, and they are not accountancy.

Tornado damage. NCEI records crop damage on 78.8% of tornado rows — better coverage than any other peril in this ranking. Tornado events reach the record through a cross-match between the SPC tornado history and NCEI's county segments, which is the one place a peril's crop figure here depends on a join rather than a direct read; the consequence is quantified in §6.

3. Method

For each event we parse the crop-damage field to dollars. NCEI writes these as strings with magnitude suffixes — 10.00M, 5K, 0.00K — so a naive numeric cast silently discards every suffixed value and returns a column that looks empty. It is worth stating because it is an easy way to conclude, wrongly, that the field is unpopulated.

We then rank perils by total recorded crop damage, and separately by dollars per damaging episode. For the headline contrast we compare drought and hail with a Mann-Whitney U test at episode level, and report Cohen's d on log₁₀-transformed values because both distributions are heavy-tailed. Perils with fewer than ~300 damaging events get bootstrap confidence intervals on the mean (10,000 resamples, seeded).

The null hypothesis we test: that the per-event crop-loss distribution is the same for every peril — that loss per event does not depend on which peril it was, and the ranking is simply a ranking of how often things happen. We test it with Kruskal-Wallis on per-event values. A chi-square on summed dollars would be invalid here: chi-square is defined on counts, and on dollars the statistic scales with the unit of measurement.

A sensitivity we build in. NCEI files one row per county or zone affected. A multi-zone episode may therefore carry a damage figure repeated across its rows. If that figure is a total allocated across zones, summing is correct; if it is one estimate copied onto each row, summing overcounts. We cannot resolve which from the record, so every total is reported twice: as recorded, and as a conservative floor that counts each uniform-value episode once.

4. Results

4.1 The ranking

Table 1 gives the full ranking under both readings; Figure 1 shows it alongside the frequency/severity structure that produces it.

#	peril	recorded	conservative floor	share	damaging events
1	drought	$29,118,680,730	$24,515,596,230	62.0%	4,662
2	freeze / cold	$7,891,281,700	$5,811,450,818	16.8%	1,224
3	hail	$4,377,658,143	$4,277,111,413	9.3%	11,377
4	wind	$2,851,393,290	$2,779,587,060	6.1%	10,582
5	heavy rain	$894,136,710	$892,633,590	1.9%	230
6	wildfire	$710,872,060	$705,827,060	1.5%	258
7	heat	$505,337,520	$504,936,520	1.1%	47
8	tornado	$364,317,960	$361,834,460	0.8%	1,868
9	winter storm	$172,846,150	$101,284,150	0.4%	284
10	flood	$42,330,000	$42,330,000	0.1%	113
11	lightning	$10,948,510	$10,932,400	<0.1%	151
12	dust	$8,100,000	$8,100,000	<0.1%	17
13	coastal	$3,199,000	$3,199,000	<0.1%	12

Table 1 — Recorded US crop damage by peril, 1993–2026. "Conservative floor" counts each uniform-value multi-zone episode once rather than summing its rows. Fog (18,300 events) and avalanche (869) are omitted from the table because both record $0; see §4.3.

The order is identical under both readings. Even at the conservative floor — which strips $4.6B from drought and $2.1B from freeze — freeze still exceeds hail. The top four hold 94.2% of the total either way.

Figure 1 — A: recorded crop damage by peril on a log scale, with the conservative floor marked as a white cut-line on each bar. B: the inversion that produces the ranking — the horizontal axis is how often a peril does recorded damage, the vertical axis is the median cost when it does, and bubble area is share of all recorded crop damage. Hail (orange) sits far right and low: the most numerous damaging peril, and among the cheapest per event.

4.2 The severity gap is the finding

Drought is not first because it happens often. It is first because of what it costs when it happens.

Hail is far more numerous. The record holds 3,579 damaging hail episodes against 550 drought episodes — a factor of 6.5 — and 11,377 damaging hail rows against 4,662 drought rows.

But the median damaging hail episode records $45,000 of crop loss against $3,127,500 for drought: a factor of 69.5. At row level the same contrast reads $15,000 against $395,000, a factor of 26 — the row figures dilute the gap, because a drought episode is spread across more rows than a hail episode is.

The distributions separate well beyond sampling noise: Mann-Whitney U = 1.58 × 10⁶ on 550 versus 3,579 episodes, p = 1.1 × 10⁻¹¹⁷, Cohen's d on log₁₀ values = 1.38 — a large effect by any convention.

Our null hypothesis — that the per-event loss distribution is the same for every peril — is rejected: Kruskal-Wallis H = 5,542, df = 12, p ≈ 0, with rank effect size ε² = 0.18.

4.3 Two perils never recorded taking a crop

Across 18,300 dense-fog events and 869 avalanches, all of them from 1996 onward, the recorded crop damage is $0.00 — not missing, recorded and zero. Both have crop-damage coverage in line with everything else (61.0% and 71.3%). They are reported here rather than dropped, because a peril that never costs a crop anything is a real finding about which hazards agriculture is exposed to.

4.4 Where the ranking is solid and where it is not

The top four are robust: each rests on thousands of damaging events, and no single event dominates. Hail's largest single entry is 2.3% of its total; drought's is 5.1%.

Below fourth place, the ranks should not be read as ordered. Heat's $505M rests on 47 events, with a bootstrap 95% CI on the mean of $357,220 to $24,257,080 — a two-order-of-magnitude interval. Dust rests on 17 events, coastal on 12. These entries establish that the peril is far behind the leaders, not that it sits at any particular rung.

4.5 The ranking survives three ways of attacking it

The result does not depend on the ambiguous episodes, on the unequal record windows, or on the exceptional years.

Uniform-value episodes. Counting each once instead of summing strips $4.6B from drought and $2.1B from freeze. The order does not change and freeze still exceeds hail (§4.1).

Unequal observation windows. Hail, wind and tornado have event records from the 1950s; the other ten perils begin in 1996. Restricting every peril to a common 1996–2025 window: drought 63.3%, freeze 17.1%, hail 8.3%, wind 5.6%. The order is unchanged and drought's share rises slightly.

Exceptional years. 2012 alone carries 20% of all recorded drought crop loss, and 2007 carries 30% of freeze. Removing them:

years removed	drought	freeze	hail	wind
none	62.0%	16.8%	9.3%	6.1%
2012	57.3%	18.8%	10.5%	6.9%
2012, 2013	53.9%	19.7%	11.5%	7.6%
2012, 2013, 2006, 2007	54.5%	16.8%	13.6%	8.8%

Removing the four costliest years in the record still leaves drought at 54.5% and hail third at 13.6%. The finding is not an artefact of the 2012 drought.

5. What this says, and what it doesn't

It says that in the measured record, agricultural loss is dominated by slow, wide, prolonged perils rather than fast, narrow, violent ones, and that the ratio is close to four to one (78.8% against 21.2%). Drought and freeze cost the crop; hail and wind cost the roof.

It does not say hail is unimportant — $4.4B is $4.4B, and hail damage is concentrated in ways drought is not. It does not say drought is worse for any particular farm. And it emphatically does not say that a drought episode is "69 times worse" than a hail episode in any physical sense: a drought record covers a zone for months, a hail record a county for minutes. The severity gap is partly a real difference in destructive reach and partly an artefact of how the two are booked. The direction is robust; the multiple is not.

6. Limitations

Damage figures are estimates, not settlements. NWS offices enter them at reporting time. They are not reconciled against insurance payouts, and practice varies by office and era.

Absent is not zero. 37% of events carry no crop-damage figure at all. If missingness correlates with severity — in either direction — every total here is biased and we cannot determine which way.

Spatial units differ by peril, systematically and in the direction that matters. Drought, freeze, heat and winter storms are recorded against forecast zones (99.9%+ zone-based). Hail is recorded against counties (99.8%). Wind is mixed (22% zone). The two heaviest perils are the zone-recorded ones. We report this rather than correct for it, because any correction would require an assumption about area equivalence that the record does not support.

Tornado loss here is a floor, not a total. The tor dex is SPC-primary, and NCEI segments that do not match an SPC tornado never reach a slot. The segment match rate is 94.0%; the unmatched segments hold a further $147.0M of crop damage. Tornadoes would stay eighth at the full figure, but no other peril in this table carries a cross-matching step, and that asymmetry should be known.

Figures are nominal, not inflation-adjusted. Dollars from 1993 and 2026 are summed as recorded. This matters less than it might: the perils are era-matched, with dollar-weighted mean years of 2007.7 for drought, 2008.5 for freeze, 2006.6 for hail and 2007.3 for wind. Adjustment would move the totals but not their order.

Report counts track observers as well as weather. Event counts rise with population and spotter density. This paper makes no trend claim for that reason: it ranks a fixed record, and does not ask whether any peril is getting worse.

7. Conclusion

Across 1.8 million recorded US storm events and $46.95 billion of crop damage recorded since 1993, the ranking is drought, freeze, hail, wind. It holds whether the ambiguous multi-zone episodes are summed or counted once, whether every peril is restricted to a common 1996–2025 window, and whether or not the four costliest years in the record are removed. Hail, the peril with an industry built around it, is third with 9.3%. The perils that take the crop arrive slowly, cover whole zones, and last for months.

---

Data and reproducibility

All data and scripts at workspaces/crop-loss-by-peril/. Source is the TerraPulse dex layer: fourteen peril Eventdexes plus flood_report, all tracing to NOAA NCEI Storm Events (ncei_storm_events), with tor additionally cross-matched to spc_tornado_history. Every figure in this paper is in data/results.json. Analysis seeded 20260829; bootstrap 10,000 resamples.

Pre-registration: the per-peril totals were computed and recorded in issue #294 before any narrative was written.

References

• NOAA National Centers for Environmental Information, Storm Events Database. https://www.ncei.noaa.gov/stormevents/
• NOAA Storm Prediction Center, Severe Weather Database (tornado history).
• TerraPulse Lab, Drought–Wildfire–AQI Cascade: Multi-Month Compound Event Chain, issue #48, 2026-04-05. workspaces/drought-wildfire-aqi-cascade-multi-month/
• TerraPulse Lab, ENSO Phase vs US Drought Severity (draft), 2026-03-19. workspaces/enso-drought-correlation/