Caribbean Flash Flood Risk Is "Extremely High" This Wet Season. AI Could Warn the Street Before the Water Does

Jacob van Ruisdael, Mountain Torrent, 1670s. The Metropolitan Museum of Art, public domain.

On July 3 the Barbados Meteorological Services extended a flash flood watch to 6 p.m. as a tropical wave sat over the island, adding up to an inch of rain to what had fallen that morning. On June 18 the same agency had issued a similar watch, warning of one to two inches of rain, water settling at the foot of hills and rising levels in ponds that are normally dry at that time of year. Each time, low-lying areas flooded or came close, and the watch lapsed once the rain moved on.

On July 4 CariCOF released its July to September outlook. It described "high to extremely high potential for flooding, flash floods, cascading hazards and associated impacts" across the Caribbean, from excessive rainfall, warming ocean temperatures and a strengthening El Nino. The outlook covers the 16 member states of the Caribbean Meteorological Organization, from Belize and Guyana to the small Eastern Caribbean islands.

What CariCOF's July outlook says

CariCOF is run by the Caribbean Institute for Meteorology and Hydrology (CIMH) in Barbados, and its quarterly outlook is the region's standard three-month climate read. The July edition lists several hazards at once:

• flash flood potential rated high to extremely high;
• frequent Saharan dust, which tends to suppress tropical cyclone formation but worsens heat and air quality;
• humid heat building into September, with recurrent heatwaves;
• severe short-term drought already present in western Cuba, Saint Lucia, and Saint Vincent and the Grenadines at the start of May.

Flood and drought in the same outlook is the practical difficulty. One island can be short of water while a tropical wave floods its neighbour, and the same country can swing between the two within a season. Any warning system built for this year has to handle both.

Why flood warnings stop at the island or parish level

The forecasts this season have been accurate. The limit is resolution. A watch for all of Barbados tells a household in a flood-prone part of St Michael the same thing it tells one on high ground in St John.

That comes from the instrument network. A dense set of river gauges, drainage sensors and automatic weather stations costs money to build and more to maintain, and most Caribbean met services have a fraction of the staff and budget of services in larger countries. Barbados is among the better-resourced services in the Eastern Caribbean and still warns at the scale its instruments support. A street-level warning without sensors behind it would be a guess, and a wrong guess costs public trust in the next warning.

The households with the least specific information are often the most exposed: people living in gullies and drainage channels because that was the land available.

Three places AI narrows the gap

Nowcasting where water will pool

Radar and satellites already estimate rainfall at regional scale. A machine learning nowcast combines those estimates with terrain, drainage, soil saturation and past flood records to predict where water will collect over the next one to six hours, at neighbourhood resolution.

This matters in a region with few gauges. A model can learn the relationship between rain intensity, slope, soil type and past flooding at the points that do have records, then apply it to gullies that have never had a sensor. Google's flood forecasting work in South Asia and Africa uses the same idea at river scale.

Landslide susceptibility maps

Landslides follow heavy rain on the steep slopes of the Eastern Caribbean and Jamaica's interior. Susceptibility models built from satellite imagery, slope angle, soil, land use and rainfall history rank hillsides by how likely they are to fail. That turns a seasonal warning into a list of slopes an emergency department can inspect before the rain and check during it.

Alerts that reach a phone in the flood zone

A nowcast helps nobody if it arrives as a press release after the water has risen. The last step is an automated SMS or WhatsApp alert to phones registered in a named flood zone, in English and in patois where that is how the community talks. With a one to six hour window, that alert is enough time to move a car, a family and important papers to higher ground.

One system for flood and drought

A flood-only model will miss the drought two parishes away. The inputs overlap: satellite soil moisture, rainfall records and terrain serve both hazards, so the same platform can flag a gully at flood risk this afternoon and a farm plot under water stress weeks before the crop visibly fails. Building one regional platform that tracks both costs less than commissioning two.

Why this is my research problem

The large climate models behind global policy need supercomputers that small island states do not have, and their output arrives too coarse for a single gully. My PhD research develops generative AI climate models intended to come close to that accuracy at a fraction of the computing cost, so a Caribbean met service can run them itself. I also chair the Caribbean AI Risk Management Council, which treats this kind of seasonal risk as part of its brief.

The weak point in my own argument is data. A nowcast trained on a handful of Barbadian gauges and a few years of flood reports will make mistakes, and nobody yet knows how many false alarms a community will tolerate before it starts ignoring the messages. The pilots below should measure that as carefully as they measure lead time.

What to do before the next tropical wave

1. National met services: list every gully, coastal road and hillside community that has flooded or slipped in the past five wet seasons, with dates and rainfall totals. This list is the training data for a nowcast or landslide model, and it can be compiled from existing records in weeks.
2. Departments of emergency management: open SMS registration for households in the listed zones now, through district emergency organisations, churches and schools, so an alert system has phone numbers ready when a model is.
3. Community groups in known flood spots: keep a flood log. Record the date, time water started rising, depth at a fixed mark and photos. Send it to the met service after each event; a year of logs from one gully is useful model data.
4. CIMH and regional partners: run one shared nowcasting and landslide pilot on two islands in the 2027 wet season, and report both lead time gained and false alarm rate, so other states can judge it on evidence.
5. Water utilities in drought-affected states: use the same soil moisture and rainfall data to set dated rationing triggers, and review them against each monthly CariCOF drought bulletin.