Seven months ago, on 28 October 2025, Hurricane Melissa made landfall near New Hope as a Category 5 storm, the strongest hurricane ever recorded striking Jamaica. At least 45 people died. Five major hospitals and more than 100 health centres were damaged. Physical damage came to US$8.8 billion, equal to 41% of Jamaica's 2024 GDP, and output fell by as much as 13% in the last quarter of 2025. The 2026 Atlantic hurricane season opens on 1 June.
On 18 May 2026 the World Meteorological Organization warned that faster-than-average sea level rise, stronger hurricanes, extreme heat and sharper swings between drought and flood are increasing across Latin America and the Caribbean. CARICOM states already lose an estimated 2% of their infrastructure capital stock each year to climate-related damage.
I have built AI systems in the Caribbean for most of my career. Several of the tools that would help this season already exist and are free or cheap, and most Caribbean agencies are not yet using them.
What Melissa cost Jamaica
The damage assessment by the World Bank and the Inter-American Development Bank put physical damage at US$8.8 billion: 41% to homes, 33% to infrastructure, 21% to other buildings and 5% to agriculture. Total damage and losses reached US$12.2 billion. Hospitals lost power and supplies when people most needed care, and hotels in the west, the country's largest source of foreign exchange, were badly hit.
Jamaica secured US$6.7 billion in international support over three years from partners including the IMF, IDB and World Bank. The economy is projected to shrink by 4.5% in fiscal year 2025/26 and return to 3.3% growth by 2027/28. World Weather Attribution found that climate change made Melissa more intense than it would have been in a pre-industrial climate.
Colorado State University expects a below-average 2026 season, with 13 named storms and 6 hurricanes, and still puts the chance of a major hurricane landfall in the Caribbean at 35%. Melissa struck in a season that was not exceptionally active.
What AI forecasting did during Melissa
Google DeepMind launched its Weather Lab cyclone model in June 2025, and the US National Hurricane Center consulted it through the season under the identifier GDMI. The NHC's review of 2025 says the model provided "very valuable guidance early on" for Melissa's track and intensity, in a storm it called "very difficult to forecast".
The University of Miami also released a tool in 2025 that automatically identifies and tracks tropical easterly waves, the disturbances off Africa that seed many Caribbean hurricanes. No automated method had existed before.
AI models such as GDMI, Huawei's Pangu-Weather and NVIDIA's FourCastNet run a global forecast in minutes, not the hours a physics model needs on a supercomputer, which makes it cheap to run many versions and see the spread. Two limits matter here. The models learn from global reanalysis data, which is less constrained by observations over the tropical oceans where Caribbean storms form. And intensity, especially rapid strengthening like Melissa's, remains harder for them than track.
Alerts people can act on
Most islands warn the public through met service bulletins on radio, television and social media. That relies on people watching the news and translating a cone map into what it means for their street.
A better system combines forecast output with river gauges, surge models and rainfall data, and sends each household a message about its own location by SMS, WhatsApp or cell broadcast. A farmer in St Elizabeth would learn that flood water is expected on her road within 18 hours and where the nearest open shelter is. A shop owner in Portmore would learn the expected surge height at his address and how long he has to move stock. Several governments built SMS and WhatsApp lists during COVID-19 that could be reused. The missing part is the software that turns model output into short, local, plain-language warnings.
Predicting damage before landfall
Machine learning models that combine a storm forecast with building data, flood maps, road networks and damage records from past storms can rank communities by expected damage. Jamaica has records from Ivan (2004), Dean (2007), Beryl (2024) and Melissa (2025) to train on.
Seventy-two hours before landfall, such a model can give ODPEM and the Jamaica Defence Force a ranked list: communities likely to need search and rescue, roads likely to be blocked, districts at risk of being cut off, and health facilities likely to lose power or water. Commanders can place people, vehicles and supplies before the storm instead of waiting for reports afterwards.
Faster damage maps and relief routing
Models trained to spot damaged roofs, flooding and blocked roads can map an island from satellite and drone imagery within a day or two of a storm, where manual surveys take weeks. That map tells responders where to go first and gives governments evidence for aid requests.
Routing software of the kind used by large logistics companies can then allocate water, food, tarpaulins and generators across dozens of communities, and recalculate as roads reopen or stocks run down. An ODPEM coordinator could be told to send a convoy to one Clarendon community first because a bridge further west will stay passable for only six more hours.
Getting farms back into production
Agriculture was 5% of Melissa's physical damage, but farm losses raised food prices and cut rural incomes for months. Drone imagery can map crop damage across a farming district within a day, giving farmers evidence for insurance and relief claims. Soil and rainfall forecasts can guide what to replant and when, and matching tools can link farmers with buyers while normal markets are disrupted.
A shared CARICOM platform
Few Caribbean states can build and maintain these systems alone. A shared CARICOM platform could hold a forecast layer checked against Caribbean storms, a common alerting system that works in English, Spanish, French, Dutch and the Creoles, damage mapping that any member can switch on after a storm, and a relief logistics tool that works across borders. The Caribbean Disaster Emergency Management Agency (CDEMA) is the obvious coordinator and the Caribbean Development Bank, with donor partners, an obvious funder. What I cannot tell you is how the platform would be paid for after the first grant runs out, and that question has sunk regional tools before.
What Caribbean governments should do this season
- National meteorological services: add AI model tracks, including Google DeepMind's publicly viewable Weather Lab forecasts, to every storm briefing alongside the NHC forecast, and log after each storm which guidance was closest.
- Disaster offices such as ODPEM: reuse the COVID-19 SMS and WhatsApp lists for location-based hurricane and flood alerts, and test them with one parish before the August to October peak.
- Telecoms regulators and operators: agree the technical and legal steps for cell broadcast emergency alerts, which reach every phone in an area without a subscriber list.
- Ministries responsible for planning and works: commission flood and surge maps for the most exposed coastlines and river basins from regional firms and universities, with a working version before peak season.
- CDEMA: prepare a costed proposal for a shared AI disaster platform, including who pays for running it after year three, and put it to CARICOM heads within 18 months.
Frequently Asked Questions
Can I see AI hurricane forecasts myself?
Yes. Google DeepMind's Weather Lab website shows its cyclone model's forecast tracks for active storms, and anyone can view them. Treat them as extra information: the official forecasts are those of the US National Hurricane Center and your national meteorological service, and evacuation orders come from your disaster office.
How often does the National Hurricane Center update its forecasts?
The NHC issues full advisories on active Atlantic storms every six hours, at 5 and 11, morning and evening, US Eastern time. When a hurricane watch or warning is in effect for land, it adds intermediate advisories every three hours. Caribbean met services issue their own local bulletins based on this guidance.
How do CCRIF payouts work after a hurricane?
The Caribbean Catastrophe Risk Insurance Facility sells parametric insurance to governments. It pays according to a model of the storm's wind, surge or rainfall over the insured country, not according to assessed damage, which is why payouts can arrive within 14 days of an event. That speed covers urgent costs, and AI damage mapping is still needed to target the rest of the recovery spending.
What is cell broadcast and does the Caribbean use it?
Cell broadcast sends an alert to every phone connected to the towers in a chosen area, without needing phone numbers or an app, and it keeps working when networks are congested. It underpins the UN's Early Warnings for All initiative, which aims to cover everyone on earth with warning systems by 2027. Several Caribbean states are working toward it, so ask your telecoms regulator what stage your country has reached.
What should a household do with a phone before a storm?
Photograph every room, your roof and your valuables before the storm, and save the pictures with scans of ID, title and insurance papers to cloud storage. Download offline maps of your area and save your disaster office's and utility's numbers. Those photos are the evidence insurers and relief programmes ask for, and they are hard to recreate afterwards.