A storm forecast is not a crop-loss forecast—WxCrop tries to bridge the gap

A hurricane forecast can trace a storm’s path days ahead. But it usually cannot tell a grower whether citrus, tomatoes or cotton are in its way—or how much of their season’s value may be at risk. In Florida, Hurricane Milton caused an estimated US$428 million in agricultural damage in October 2024; the preliminary estimate arrived ten weeks later, and the final one about ten months after landfall. WxCrop, a new system from Minghan Yu, Mengyu Dai, Yixin Wen and Haizhao Yang, aims to move part of that answer closer to the moment when people can still act (Yu et al., 2026).

The system joins weather observations and forecasts to crop maps, growing calendars and insurance-loss records. It estimates which crop groups face risk, where, and what fraction of expected season value might be lost. In the paper’s Florida case study, the hurricane index detected recorded cyclone days with an AUC of 0.98, a measure of how well a score separates one class from another. The authors also report that held-out losses were reproduced to within 2%, and that independent production records supported the model’s county rankings, with a rank correlation of 0.63.

That does not make a forecast a field inspection. WxCrop’s estimates are advisory and county-level, calibrated against insured losses rather than confirmed damage on every parcel. But the central ambition is important: turn a forecast of dangerous weather into a sourced, crop-aware estimate of what that weather may mean for livelihoods and food production.

The Science

Agricultural disaster costs are large, but their timing and geography vary. The paper cites an estimate of US$3.26 trillion in global agricultural losses from disasters between 1991 and 2023—roughly 4% of global agricultural GDP per year—with climate-related hazards accounting for most of that total. Early warnings can help people prepare, but the warning has to answer a practical question: which assets are exposed, and how much is at stake?

That question is difficult because the evidence lives in separate systems. A forecast may describe wind and rainfall; a crop inventory may show what is planted; a growing calendar says whether the crop is newly planted or nearly ready to harvest; insurance records document payouts later. Those datasets use different maps, dates and units. Joining them repeatedly for a particular county and day takes time and technical skill.

WxCrop builds a common route through those records. A user selects a hazard and date, or asks in ordinary language. The map displays relevant weather layers alongside agricultural parcels, crop types, and estimated growth stages. For hurricanes, the system uses wind and seven-day rainfall. It draws tropical-cyclone winds from the National Hurricane Center’s best-track archive, which reconstructs storms from observations, rather than relying on gridded weather products that can smooth out a storm’s strongest winds. For other days, it uses gridded analyses. The system labels values as observed, analysed or forecast and identifies their sources.

The crop map matters because exposure is not the same as vulnerability. A field planted with an annual crop may have only a fraction of its expected season value invested early in the season; as harvest approaches, more value is at risk. A citrus grove carries the value of its trees year-round, while the fruit’s progress through the season changes the harvest value exposed. WxCrop uses growing calendars and cost information to estimate that date-specific vulnerability.

The workflow then produces a weather-only hazard index and, where the index is high enough, an estimate of expected loss by crop group. The language model is not asked to invent a damage number. Instead, defined software tools calculate and retrieve the values; the conversational interface explains them and gives their sources. That distinction is part of the design: an answer should be traceable to a weather record, crop inventory and model output.

Figure 1 shows the interface for Hurricane Ian’s Florida landfall on 28 September 2022: weather layers, county and parcel crop information, growth stages, and a question about wind at a citrus grove. The example illustrates the system’s intended scale transitions—from a statewide view to a particular mapped parcel—while retaining the caveat that a wind value for a grid cell is not a direct measurement at the tree.

What They Found

For hurricanes, the index combines wind and rain. The reasoning is physical: strong winds can topple trees, and saturated soil can weaken their anchoring. The authors tested candidate weather factors against insured-loss records using a pre-specified causal-analysis procedure. Wind and rainfall showed useful signals, as did their interaction; other candidates, including several hydrological and land characteristics, either repeated those signals, lacked useful data or did not show a comparable relationship in the tests reported.

The resulting hurricane index, written as a wind load multiplied by a rainfall amplifier, is designed to rise as the weather becomes more damaging. The wind component grows with the square of sustained wind speed, reflecting the way dynamic pressure rises with wind speed. It assigns each county-day to one of four tiers: baseline, elevated, likely or severe. The boundaries are fixed from the historical distribution of index values during exposed conditions, rather than chosen anew for each storm.

The index’s tiers track insurance outcomes. In county-months in the lowest tier, 16% had insured crop losses, about as often as months with no storm, according to the paper. In the highest tier, 81% had insured losses and 66% had losses above US$100,000. These are historical associations, not a promise that a particular field will be damaged. But they give the tiers an interpretable meaning: they summarize how often similar hazard conditions have been followed by a recorded loss.

The hurricane index separates county-months by insured losses

Label	Value
Lowest tier: any insured loss	16 %
Highest tier: any insured loss	81 %
Highest tier: loss above $100,000	66 %

• Lowest tier: any insured loss 16 %
• Highest tier: any insured loss 81 %
• Highest tier: loss above $100,000 66 %

The paper reports 16% in the lowest tier, 81% with any insured loss in the highest tier, and 66% in the highest tier with losses above $100,000.

The contrast can be seen in the Ian example. DeSoto County, near the storm’s track, received a severe-tier value of 27.9; Jackson County in the Panhandle received a baseline value of 0.20. The county wind statistic is based on the 90th percentile of storm-wind values across county grid cells, so it is not interchangeable with the peak wind estimated for a specific parcel.

For a high-tier county, the loss model combines the hazard index with crop susceptibility and the share of the crop’s season yield still in the field. It estimates both the expected loss if damage occurs and the probability of an insured payout, then multiplies the two. In the example for DeSoto County’s citrus, the model estimates an 83% loss fraction if damage occurs and a 98% payout probability, resulting in an expected loss of 82% of the crop group’s expected season value. That is an estimate for the county’s insured units, not a claim that every grove loses that share.

WxCrop also covers freezes and dry heat. Its freeze index uses the minimum temperature and night-time cloud cover; clear skies can increase radiative cooling, the loss of heat from the ground and plants to the atmosphere. Its dry-heat index combines a long-term precipitation deficit relative to evapotranspiration—the water lost from land and plants—with accumulated heat above 32°C. Each hazard has its own weather inputs and calibrated loss model. The paper’s interface shows example outputs for a freeze in Hendry County and a dry-heat episode in Jackson County, alongside the hurricane case.

The authors report three headline checks: hurricane-day detection at AUC 0.98, held-out recorded losses reproduced within 2%, and county rankings that correlate at 0.63 with independent production records. These measures test different things. Detection asks whether the index identifies cyclone days; reproducing held-out losses tests estimates against insurance records not used in the same way for fitting; the independent comparison asks whether counties with larger estimated impacts also rank higher in production-based evidence. Together they are more informative than a single accuracy score, though they do not establish that every forecast estimate will match later field damage.

The underlying payout pattern also helps explain why a Florida tool cannot simply be exported unchanged. In the paper’s table, hurricanes and tropical cyclones account for 36.6% of weather-caused payouts in Florida, while drought and heat account for 13.1%. Across nine south-eastern states, drought and heat make up 59.3%, and hurricanes 5.2%. The same region can face a different dominant risk profile just across a state boundary.

Weather-related insured payouts have different profiles across Florida and the Southeast

Label	Value
Hurricane / tropical	36.6 %
Excess moisture / flood	30.2 %
Freeze / cold	14.2 %
Drought / heat	13.1 %
Wind (non-cyclone)	3.4 %
Hail	0.9 %
Other weather	1.6 %

• Hurricane / tropical 36.6 %
• Excess moisture / flood 30.2 %
• Freeze / cold 14.2 %
• Drought / heat 13.1 %
• Wind (non-cyclone) 3.4 %
• Hail 0.9 %
• Other weather 1.6 %

Shares of weather-caused payouts; total payout amounts reported are $1.4 billion for Florida and $34.2 billion for the nine-state region.

Why This Changes Things

The useful shift is from predicting weather to estimating consequences. A storm track is a hazard forecast; a crop-aware impact estimate adds the question of what is planted, how far through the season it is, and what prior losses under similar conditions look like. That extra layer can help an emergency manager decide where to send inspectors or help a grower compare scenarios before committing to planting, harvesting or other time-sensitive choices.

Timing matters because agricultural assistance often follows verification. The paper notes that federal crop-insurance payouts depend on loss verification for insured units, while a county disaster designation requires documented production losses of at least 30% for one crop. Florida’s recent hurricane damage estimates have taken months to finalize, and some recovery programmes take years to reach applicants. WxCrop cannot remove the need to verify claims, but an early map of likely exposure could help agencies prioritize where verification and outreach begin.

The system’s design addresses another familiar weakness of AI assistants: fluent but unsupported answers. A general-purpose language model can explain information, but it is not a dependable calculator or source of truth. WxCrop makes the numeric estimate in a separate calibrated model and uses the language model to retrieve and explain the result. The source, units and data version remain visible. In principle, this makes the tool easier to audit and contest than a prediction delivered as an unexplained sentence.

There is also a potentially valuable research use. A common workflow can let extension specialists and scientists repeatedly compare weather, crop stage and outcomes without rebuilding the data pipeline for every event. The authors’ goal is an early-warning layer that can be extended to other hazards and regions when comparable weather, crop and loss records are available. The Florida demonstration is a starting point, not evidence that the model already works everywhere.

What’s Next

The central limitation is the target the model learns to predict: insured losses. Insurance records are valuable because they are extensive and tied to crop, place, month and reported cause. But payouts reflect more than physical damage. Coverage, policy terms, reporting and insurance participation influence what appears in the records. An uninsured farm, an underreported loss or damage that does not trigger a payout may be missing from the signal. The estimates should therefore guide attention, not determine compensation or substitute for inspections.

Geography and resolution matter too. The loss outputs are county-level estimates for crop groups, calibrated at the scale of insured units. A county average can hide substantial differences between farms, soils, varieties and management practices. Parcel maps make the interface look field-specific, but the paper distinguishes parcel-level crop and exposure information from the coarser scale at which the loss model is calibrated. Users should not mistake a detailed map for equally detailed damage certainty.

Forecasts add another layer of uncertainty. WxCrop can use forecast guidance for the coming days, but a projected wind field or rainfall total can change as a storm evolves. The paper appropriately labels forecasts and distinguishes them from observed or analysed weather. In practice, an estimate made several days before landfall should be revisited as new forecast data arrive and treated as a scenario, not a final bill.

Further tests should establish how well the approach transfers beyond Florida and beyond the three demonstrated hazards. Hurricanes, freezes and dry heat leave different signatures in insurance records; other risks, including flood and compound events, may demand new indices, better crop data or different loss models. Independent field surveys, production statistics and post-event assessments can clarify where estimates are reliable, where they systematically miss damage, and whether early use actually speeds aid or improves decisions.

That is the promise—and the standard WxCrop must meet. Forecasts already tell us when dangerous weather may arrive. A useful agricultural warning must also tell people what is growing in its path, what value is exposed, and how uncertain that estimate remains. If the system can keep those answers fast, transparent and calibrated as it expands, it could turn an early warning from a description of the sky into a practical tool for protecting farms and the communities that depend on them.