From extreme weather to agricultural consequences with agentic AI

Minghan Yu; Mengyu Dai; Yixin Wen; Haizhao Yang

physics.ao-phOct 29, 2026

From extreme weather to agricultural consequences with agentic AI

Authors: Minghan Yu, Mengyu Dai, Yixin Wen, Haizhao Yang

Organizations: Department of Mathematics, University of Maryland, College Park, College Park, MD, USA. · Stony Brook University, Stony Brook, NY, USA. · School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY, USA. · Department of Computer Science, University of Maryland, College Park, College Park, MD, USA.

Abstract

Extreme weather is forecast days ahead, yet its consequences for agriculture remain hard to predict: a forecast says where a storm will strike, not which crops will suffer or how much is at risk, and the records that could tell span many sources. We present WxCrop, the first agent that judges agricultural hazards from physically interpretable weather indices examined by causal analysis, and the first that combines weather and agricultural information to predict crop loss reliably for past and forecast dates. We demonstrate it in Florida on hurricanes, freezes and dry heat, with a loss model calibrated on insurance records for each, and the same procedure extends to other hazards and areas. Three checks show reliability: the hurricane index detects recorded cyclone days with an AUC of 0.98, recorded losses in held-out years are reproduced to within 2%, and independent production records confirm the county rankings (rank correlation 0.63). Users ask the agent for each crop group's loss as a fraction of expected season value, every number sourced, so that governments, growers and scientists respond to agricultural disasters quickly.

Figures & tables

Figure 2: One number for the hazard. All panels are for 28 September 2022. a , The hurricane hazard index $H$ on its native grid. b , The county-level layer for the same index, with the hover card of DeSoto County. c,d , The assistant’s answers for DeSoto County, near the storm’s track (severe tier), and for Jackson County, in the Panhandle (baseline tier). e , The proportion of county-months with insured weather losses (blue) and with losses above $100,000 (orange) at each level of $H$ , with the four tiers shaded. — Figure 2: One number for the hazard. All panels are for 28 September 2022. a , The hurricane hazard index $H$ on its native grid. b , The county-level layer for the same index, with the hover card of DeSoto County. c,d , The assistant’s answers for DeSoto County, near the storm’s track (severe tier), and for Jackson County, in the Panhandle (baseline tier). e , The proportion of county-months with insured weather losses (blue) and with losses above $100,000 (orange) at each level of$ H$ , with the four tiers shaded.

Figure 3: From hazard to a percentage of the crop. a , A county in the lowest hazard tier (Jackson County, hurricane, 28 September 2022): the answer ends at the tier. b , A county in the severe tier on the same day (DeSoto County): expected loss by crop group. c,d , The same output for a freeze (Hendry County, 11 January 2010) and for dry heat (Jackson County, 15 August 2011). e , The county-level layer for citrus on the same day, with the hover card of DeSoto County.

Figure 4: How far the answers can be trusted. a , Receiver operating characteristic curve of the hurricane index

H

against recorded tropical-cyclone days. b , Proportion of winter county-months with insured freeze losses (blue) and with losses above

100,000 (orange) in each range of the freeze index, with the number of county-months below the axis. c , Proportion of insured field-crop county-seasons with a claim above

100,000 (orange) and of seasons with severe drought in the U.S. Drought Monitor (green) in each tier of the dry-heat index. d , Hurricane predictions against recorded insured losses, summed over storm county-months with each year held out. 1 means the totals agree and the grey band is a factor of two. e , Rank of the predicted citrus loss against the rank of the recorded citrus production decline, for 110 county-seasons, with the 11 counties of the Ian season in orange. f , Calibration of the probability of an insured loss, in ten equal groups of storm county-months.

Hazard family	Florida (%)	Nine south-eastern states (%)
Hurricane / tropical ∗	36.6	5.2
Excess moisture / flood	30.2	21.4
Freeze / cold ∗	14.2	3.7
Drought / heat ∗	13.1	59.3
Wind (non-cyclone)	3.4	4.2
Hail	0.9	5.9

Table 1: Percentage of weather-caused crop-insurance payouts by hazard family, 2004–2024. USDA Risk Management Agency Cause of Loss records [ 8 ] , commodity years 2004–2024, payouts with a weather cause. Families group the recorded causes: hurricane/tropical (hurricane, tropical storm, tropical depression); excess moisture/flood (excess moisture, flood, cold wet weather); freeze/cold (freeze, frost, cold winter); drought/heat (drought, heat, hot wind, failure of irrigation supply). Non-weather causes (price decline, wildlife, disease, insects, mycotoxin) are excluded from the denominator. The nine states are Alabama, Florida, Georgia, Louisiana, Mississippi, North Carolina, South Carolina, Texas and Virginia. The three families treated in this paper are marked with an asterisk.

Kind	Quantity or use	Dataset	Provides	Period
Weather	Wind in cyclones	HURDAT2 best track [ 43 ]	Centre, maximum wind and radii, six-hourly	2004–2024
	Wind elsewhere	ERA5 [ 44 ] , High-Resolution Rapid Refresh (HRRR) [ 45 ]	Daily maximum 10-m gust	2004–2024, HRRR from 2014
	Threatening cyclone	National Hurricane Center forecast	Forecast track and wind radii	Up to five days
	Rain	Analysis of Record for Calibration (AORC) [ 46 ]	Daily precipitation (1 km)	2004–2024
	Freeze	Parameter-elevation Regressions on Independent Slopes Model (PRISM) [ 47 ] , ERA5	Minimum temperature, night cloud cover	Winters 2003–26
	Drought and heat	gridded Surface Meteorological dataset (gridMET) [ 48 ]	Precipitation, evapotranspiration, maximum temperature	2003–2026

Table 2: Data used in this study. Each row gives the quantity, the dataset that supplies it, what it provides and the period used.

Step	Candidate (data)	Result	Decision
Inputs (area under the curve, AUC)	Wind, rain, ponding depth, wet hours (HURDAT2, AORC, National Water Model)	AUC for cyclone days: wind 0.980, rain 0.871, ponding 0.838, wet hours 0.729	Wind and rain
Form (area under the curve, AUC)	Wind alone above a zero point, wind $\times$ rain, wind $+$ rain dose, exponent 2 or 3, zero point fixed or free	AUC for cyclone days 0.970, 0.972, 0.921, best 0.981. Agreement with UF/IFAS reports 0.689, 0.703, 0.703, best 0.704 (exponent 2, free zero point)	Dynamic pressure, free zero point
Design (difference-in-differences, DiD)	$H\geq 2$ , Florida, 24,053 rows	Effect $+0.165$ [0.120, 0.210]. Before the storm $-0.016$ , $-0.002$ , $+0.009$ . Tiers $+0.008$ , $+0.037$ , $+0.104$ , $+0.287$ . Model $+0.179$	Passed
Factor (permutation)	Rain $\times$ wind (AORC)	$+0.066$ [0.030, 0.102], $p<0.005$	Adopted, multiplier
	Ponding, wet hours (National Water Model)	$+0.017$ , $+0.006$ , $p<0.005$ , small and tied to rain	Set aside
	Water-table depth, drainage class, tree age, days since storm	$p=0.27$ –0.98, or sign opposite	Set aside

Table 3: Hurricane index

H

: tests, candidates and decisions. Effects are differences in the damage fraction [95% cluster interval], defined in Methods.

p

is the proportion of 200 within-county permutations that reached the observed value (below 0.005 when none did).

Step	Candidate (data)	Result	Decision
Form (area under the curve, AUC)	Depth below 0 ∘ C, below the second or fifth percentile design temperature (ERA5, PRISM)	AUC for freeze days 0.870, 0.935, 0.981. PRISM farmland-weighted mean 0.991	PRISM, fifth percentile
Design (difference-in-differences, DiD)	$F_{z}\geq 4.8$ ∘ C-equivalent, Florida winter, 10,099 rows	Effect $+0.046$ [0.009, 0.083]. Before the freeze $-0.008$ , $+0.001$ , $+0.003$ . Tiers $+0.008$ , $+0.024$ , $+0.061$ , $+0.087$ . Model $+0.053$	Passed
Factor (permutation)	Night cloud cover	$c=-0.243$ , $p<0.005$ , sign as expected. Rank correlation 0.334 against 0.327	Adopted, clear-sky factor
	Freezing degree-hours †	$c=+0.23$ , $p<0.005$ , sign as expected. Rank correlation 0.366 against 0.364, correlated 0.82 with depth	Verified, not deployed
	Night wind, previous 14-day temperature	$c=-0.262$ ( $p<0.005$ ), $c=-0.110$ ( $p=0.17$ ), sign opposite	Set aside
	Previous-day maximum † , previous 30-day cold † , soil water	$p=0.21$ , 0.53, 0.90	Set aside

Table 4: Freeze index

F_{z}

: tests, candidates and decisions. Effects are differences in the damage fraction [95% cluster interval], defined in Methods.

p

is the proportion of 200 within-county permutations that reached the observed value (below 0.005 when none did). For factors,

c

is the exponent of a multiplier

\exp(cz)

on the index. Factors marked † need hourly fields and were tested on the ERA5 index, the others on the PRISM index.

Step	Candidate (data)	Result	Decision
Form (area under the curve, AUC)	Deficit window of 30 to 270 days, precipitation alone (gridMET, SPEI)	AUC for severe drought 0.71, 0.81, 0.85, 0.88, 0.89, 0.85. Precipitation alone 0.83	120 days
Design (difference-in-differences, DiD)	Season index classes, Florida insured field crops, 1,125 rows	Effect of the top decile $+0.112$ . Classes $+0.008$ , $+0.051$ , $+0.064$ , $+0.112$ . Model $+0.054$ (mean effect $+0.028$ [ $-0.009$ , 0.066])	Dose response monotone
Factor (rank correlation, permutation)	Flash-drought term (30-day deficit)	Rank correlation 0.577 against 0.545 for the slow deficit alone	Adopted, added to the dose
	Heat (degree-days above 32 ∘ C)	$c=0.117$ against a permutation 95th percentile of 0.072	Adopted, multiplier
	Vapour-pressure deficit, antecedent rain, irrigated fraction, soil water capacity	$p=0.14$ , 0.40, 0.41, sign opposite. Antecedent rain depends on the window	Set aside

Table 5: Dry-heat index

D_{z}

: tests, candidates and decisions. Effects are differences in the damage fraction [95% cluster interval], defined in Methods.

p

is the proportion of 200 within-county permutations that reached the observed value (below 0.005 when none did). For factors,

c

is the exponent of a multiplier

\exp(cz)

on the index.

Step	Candidate (data)	Result	Decision
Label	Damaged acreage, loss cost, recovered damage fraction	Rank correlation with county citrus production shortfalls 0.36, 0.47, 0.58	Recovered damage fraction
Crop dependence	Effect of $H\geq 2$ by crop group (difference-in-differences, DiD)	Citrus $+0.30$ , cotton $+0.32$ , peanuts $+0.16$ , tobacco $+0.13$ , vegetables $+0.06$ , grains and oilseeds $+0.04$ , other fruit $+0.02$ , sugarcane $-0.06$	Susceptibility $\theta_{g}$ by group
Structure	Four equally weighted factors, wind alone, $H$ with $\theta_{g}$	Storm-month rank correlation, leave-one-year-out, 0.351, 0.361, 0.397	$S_{g}=1-\exp(-\theta_{g}\varphi_{g}^{\gamma}H)$
Yield vulnerability	Floor of 0.02, establishment-cost fractions	Rank correlation 0.405, 0.405	Establishment-cost fractions
Storm periphery, moisture	Fraction of county $\geq 18$ m s -1 , thresholds by group, hours above 18 m s -1 , rainfall extremes, wet-day runs, soil moisture, ponding	Every metric within the permutation spread	Set aside
Payout probability	$P_{g}$ with a background rate and a storm scale	Brier score 0.177 against 0.224 for the climatological rate	Adopted

Table 6: Loss model: tests, candidates and decisions. Rank correlations are Spearman correlations. Effects of

H\geq 2

are design estimates on the damage fraction by crop group.

Disaster Damage Assessment

Disaster damage assessment uses imagery and geospatial data to locate, classify, and quantify damage for emergency response and recovery.

See all 51 papers

Agent-Based Post-Hoc Correction of Agricultural Yield Forecasts
May 12, 2026Matthew Beddows, Aiden Durrant, Georgios LeontidisTime Series Forecasting LLM Agent Orchestration
School of Natural and Computing Sciences University of Aberdeen, UK · School of Computing Sciences University of East Anglia, UK · Department of Physics and Technology UiT The Arctic University of Norway
Agri-SCHOLAR: Simulation-Grounded In-Season Rolling Advisories for Agricultural Crops Using LLM-Bounded Optimization
Oct 22, 2026Geetha Charan, Vedant Balasubramaniam, Manojkumar Patil +5AI-Assisted Decision Making Precision Agriculture
Indian Institute of Science, Bengaluru, India · University of Michigan, Ann Arbor, USA
SIREN: Towards End-to-End Extreme-Weather Early Warning with Experience-Grounded LLM Agents
Jul 27, 2026Hang Ni, Weijia Zhang, Fan Liu +2AI Agent Benchmarks Extreme Weather Forecasting
Artificial Intelligence Thrust The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Department of Civil and Environmental Engineering The Hong Kong University of Science and Technology Hong Kong, China · Otto Poon Center for Climate Resilience and Sustainability, The World Sustainable Development Institute.

May 12, 2026J/K move · Enter open · S save

Agent-Based Post-Hoc Correction of Agricultural Yield Forecasts

Matthew Beddows, Aiden Durrant, Georgios Leontidis

School of Natural and Computing Sciences University of Aberdeen, UK · School of Computing Sciences University of East Anglia, UK · Department of Physics and Technology UiT The Arctic University of Norway

Time Series Forecasting LLM Agent Orchestration LLM-Based Time Series Forecasting Crop Yield Prediction

Open paper AskPDF arXiv

Accurate crop yield forecasting in commercial soft fruit production is constrained by the data available in typical commercial farm records, which lack the sensor networks, satellite imagery, and high-resolution meteorological inputs that most state-of-the-art approaches assume. We propose a structured LLM agent framework that performs post-hoc correction of existing model predictions, encoding agricultural domain knowledge across tools for phase detection, bias learning, and range validation. Evaluated on a proprietary strawberry yield dataset and a public USDA corn harvest dataset, agent refinement of XGBoost reduced MAE by 20% and MASE by 56% on strawberry, with consistent improvements across Moirai2 (MAE 24%, MASE 22%) and Random Forest (MAE 28%, MASE 66%) baselines. Using Llama 3.1 8B as the agent produced the strongest corrections across all configurations; LLaVA 13B showed inconsistent gains, highlighting sensitivity to the choice of refinement model.

Abstract

Disaster Damage Assessment

Explore similar work

Agent-Based Post-Hoc Correction of Agricultural Yield Forecasts

Agri-SCHOLAR: Simulation-Grounded In-Season Rolling Advisories for Agricultural Crops Using LLM-Bounded Optimization

SIREN: Towards End-to-End Extreme-Weather Early Warning with Experience-Grounded LLM Agents