Curaçao's 70,000-Page Weather Archive and What AI Can Do With It

TL;DR

On 28 September 2026 the International Panel on Deltas and Coastal Areas announced that the historical weather archives of Aruba, Curaçao and Sint Maarten are digitised, going back to 1880. The regional outlook for September to November puts the ABC islands among the areas with the highest chance of rainfall deficits, and NOAA gives a 75% chance that this El Niño will be the strongest since 1950. AI can read the old pages, check them and turn them into drought and heat warnings, and the first step this week is a 14-day accuracy test on 50 pages.

What happened

The announcement, published by 721news on 28 September 2026, says data rescue and digitisation are finished for the historical meteorological archives of Aruba, Curaçao and Sint Maarten, with records for Bonaire, Sint Eustatius and Saba included. The material comes from an archive of about 70,000 pages inherited from the former Netherlands Antilles, reaches back to 1880, and includes the 1991 to 2020 climate reference period. The partners named are the Departamento Meteorológico Aruba, the Meteorological Department Curaçao, the Meteorological Department Sint Maarten and the Royal Netherlands Meteorological Institute (KNMI). The stated aims are sharper climate projections for 2050 to 2100, better early warning and better adaptation planning.

The announcement leaves out what a data officer would ask first. It does not say how many of the values have been keyed in as numbers, how many have been checked, or what format the data sits in. It says access followed legal review and intellectual property agreements, and it does not publish the terms for outside users such as a researcher or a company. The scanned pages exist. Whether the numbers on them are a usable dataset is a separate question, and this article treats it as open.

The timing matters because the season is unusual. The Caribbean Institute for Meteorology and Hydrology (CIMH) published its September to November 2026 climate messages, which say seasonal rainfall is likely to be unusually low at the peak of the wet season, with the highest chances of rainfall deficits in Hispaniola, Puerto Rico, the US Virgin Islands, the Lesser Antilles and the ABC islands. They also point to unusually warm days and nights, more humidity and longer hot spells.

The numbers

The cause is El Niño. In its 10 September 2026 discussion, NOAA's Climate Prediction Center reported sea surface temperature anomalies above +3.0°C in the eastern equatorial Pacific, a greater than 90% chance of what it calls a "very strong" event through the Northern Hemisphere winter, and a 75% chance of a historic event in October to December that would exceed every El Niño since 1950. NOAA adds that impacts are more likely with an event this size but not guaranteed. The next update is due on 8 October.

Reading paper is the slow part of data rescue, and software is getting better at it. MeteoSaver, an open-source tool described in Geoscientific Model Development in April 2026, found daily temperature values on 95% to 100% of the cells in the sheets it was tested on, and, of the values it could confirm, a median of 74% matched the manual transcription within a 0.2°C margin, with a median absolute error of 0.3°C. The test covered 10 sheets from the Democratic Republic of Congo. The quality checks set aside 25.6% of the values as unconfirmed, and the authors note that regional thresholds can screen out real extremes.

Water is why this matters on an island with little of it. Technical reporting on the Santa Barbara plant puts average rainfall at about 500 mm a year, the only natural freshwater on the island, and says the plant provides drinking water for about half the population.

The people and the money are on the same island. Curaçao's Central Bureau of Statistics counted 156,115 residents on 1 January 2025. In 2025 the island received 788,427 stay-over visitors, up from 700,242 in 2024, and the CBS put economic growth at 5.1%, with accommodation and food services up 10.8%. We found no published figure for how much water those visitors use, and that gap is part of the case for measuring.

What this means in Curaçao

I apply the method I call People First, Problem Second, AI Last. A test we suggest: write the human need and the problem in two sentences each without using the word AI. The need: a farmer in the west of the island, a hotel engineer in Pietermaai and a ward nurse in Willemstad each have to decide this month whether this dry spell is normal for Curaçao or a departure from it. The problem: normal is defined by a record, the record is 70,000 pages across several islands, and nobody can query a page.

Neither sentence needs AI. The need is a trusted answer to "how unusual is this?" and the problem is that the record is not yet something a person can ask. AI comes last, as a tool for the one step that is slow, which is turning the pages into numbers that can be checked and then used.

The announcement says the 1991 to 2020 reference period is covered, which gives a model a recent baseline, and we expect the departments that keep the records to be its natural users. The island also has a clear customer for the output. Aqualectra signed a 20-year agreement with Seven Seas Water in November 2024 for a new desalination plant at Hato West, with operations scheduled for October 2026, and a utility that is about to bring a plant online wants to know how dry the year ahead will be. I have not found confirmation that the plant is running.

One caution on the evidence. The tests quoted above were run on records from elsewhere. The study covered sheets from Congo, and we expect Dutch Caribbean pages, which were written by different hands in different decades, to behave differently. No published test exists on this archive, and our expectation is an inference, not a result.

Four jobs AI can do with a digitised weather record

1. Read the pages

Handwriting and table recognition is the first job, and it is where the MeteoSaver figures apply. A tool that confirms about three values in four and sets the rest aside leaves a person to read the set-aside values and to sample the confirmed ones. That is still a saving on 70,000 pages, provided someone budgets the checking time and does not treat the machine output as final.

2. Check the values against physics

A transcribed number can be wrong and still look plausible. Researchers at KNMI and Radboud University reduce this by testing outputs against physical relationships between variables, so, for example, a daily minimum that exceeds the daily maximum is flagged for review. We expect the same rule to run across all three islands quickly once the values are keyed in, though no one has timed it on this archive. Software finds the impossible entries quickly. It cannot tell whether a plausible but wrong 31°C was meant to be 29°C, and that is why the human sample stays.

3. Say how unusual this season is

With a clean record, an analyst can measure how far this season sits from the 1991 to 2020 normal for each island and publish the answer weekly in plain words. Statistical models can then estimate how often a season this dry has happened before. The answer is more useful to a farmer than a seasonal outlook for the whole Caribbean, because it is about Curaçao's own gauges.

4. Warn people in their own language

CIMH describes drought as evolving in the ABC islands, with high confidence, and lists heat stress as a particular concern in the Lesser Antilles and the Guianas. A heat alert for Curaçao is our inference from the warm outlook, and a drought notice is the safer first product. A warning is only useful if it arrives in a language the person reads, and Curaçao's official languages are Papiamentu, Dutch and English. Software can draft the message and send it by SMS, WhatsApp and radio script as soon as a threshold is crossed. A Papiamentu speaker should write and check the first version, because a translation tool trained mostly on other languages will state its mistakes with confidence.

The same logic reaches Aruba and Sint Maarten, which share the archive project and face a different mix of risks. Sint Maarten sits in the hurricane belt, and a clean record there helps with storms and rainfall extremes more than drought.

What a ministry or funder should ask for

A digitised archive is an asset only if someone can use it, and the people who fund these projects can set the terms at the start. Four requests cost little and change what is delivered. The first is a published error rate, taken from a human-checked random sample of the transcribed values, so that every user knows how far to trust the dataset. The second is a licence and a named custodian: one office in Curaçao that answers for the data, takes corrections and decides who may use it.

The third is a Curaçao accuracy figure from any vendor, university or foreign partner that offers a model, because a score on records from Congo or the Netherlands says little about pages written on these islands. The fourth is a delivery date for a plain-language weather bulletin in Papiamentu, Dutch and English, issued weekly during the dry season. A bulletin forces the data into the hands of the people who decide about water, and it shows within a month whether the archive is working.

What to do this week

These steps are written for three readers: a records officer at a meteorological department, a hotel or school facilities manager, and a farmer with a rain gauge. The test in the first step is one we suggest, and it is not a method Adrian has published.

1. Test the transcription. Pull 50 scanned pages at random from the digitised set, or from your own paper logs, have two people key the values by hand, run one AI transcription tool over the same pages, and compare. Record the share of values that match and the share that fail a physical check. Finish within 14 days and publish the error rate with the dataset.
2. Ask the Meteorological Department Curaçao in writing which stations and years are in the digitised set, in what format, and under what licence a business or school may use them.
3. Measure your own baseline. Read your water meter and your rain gauge at the same hour every day for 14 days, change one thing (irrigate at dawn instead of midday, or fix the loudest leak), and read for another 14 days. Compare litres per day. A model needs a record, and one dry season is one data point.

Where this can go wrong

The first harm is a bad baseline. If transcription errors enter the record unflagged, every later drought estimate inherits them, and the people who pay are the farmers, insurers and utility planners who act on the numbers. The guardrail is a published error rate from a human-checked sample, plus physical consistency checks on every value.

The second harm is about who benefits. The archive is shared with a foreign institute, and models built on it could end up as a product that the islands buy back. I know of no documented case of this on these islands, and the partners include the islands' own meteorological departments, so treat it as our inference and a risk to manage. The pattern I call Extracted Intelligence is data leaving as raw material and returning as something the community has to pay for. The guardrail is to settle licence terms in writing now, and to ask for the same access to derived products that the foreign partners have.

The third is Synthetic Confidence Risk. A model trained mostly on other regions can state drought odds for a small dry island with more authority than the record supports. In the KNMI abstract, a text-recognition model fine-tuned on training data generated from the pipeline's own outputs rose from 19% to 81% on unseen data. Our inference is that accuracy depends heavily on how well the training data fits the pages, and no one has tested a model on Dutch Caribbean handwriting. The guardrail is to ask any vendor for accuracy figures on a Curaçao sample, and to treat the absence of a figure as unknown accuracy.

The weakest part of this argument

The case rests on a press report that does not say what the 70,000 pages contain as numbers. The accuracy figures come from tests in Congo and from a conference abstract, and no one has published a test on Dutch Caribbean handwriting. I have also not confirmed that the Hato West plant is operating. If the archive turns out to be scanned images with only a small share keyed in, the first job is the dull one, and it will take longer than a season.

Questions readers ask

Where does Curaçao's drinking water come from?

Rain averages about 500 mm a year and is the only natural freshwater, so desalination fills the gap. The Santa Barbara plant provides drinking water for about half the population, according to technical reporting on the plant. Aqualectra's 20-year contract with Seven Seas Water, announced in November 2024, calls for at least 25,000 cubic metres a day from a new plant at Hato West.

Can AI read old handwritten weather records accurately?

Accuracy depends on the method and the pages. A KNMI and Radboud University team reported up to 97% precision from guided multimodal language models with strict filtering, though that figure holds only for values that pass strict filtering, at the cost of a lower inclusion rate that the abstract does not quantify. Fine-tuning a text-recognition model lifted its accuracy from 19% to 81% on unseen pages. These figures come from a conference abstract, so ask any vendor for a result on your own records.

Will El Niño make Curaçao wetter or drier?

The CIMH climate messages for September to November 2026 list the ABC islands among the places with the highest chances of rainfall deficits, with Hispaniola, Puerto Rico, the US Virgin Islands and the Lesser Antilles, and describe drought as evolving in the ABC islands with high confidence. Wet-season conditions can still raise flood potential, so plan for drought with a flash-flood plan alongside it.

Who owns the digitised weather data, and can a Curaçao company use it?

The announcement names the Aruba, Curaçao and Sint Maarten meteorological departments and KNMI as partners, and says access followed legal review and intellectual property agreements. It does not publish terms for outside users. Until it does, treat the data as not yet open. Write to the Meteorological Department Curaçao and ask for the station list, the file format and the terms of use before you build a product on it.

Which languages should heat and drought alerts use in Curaçao?

Curaçao's official languages are Papiamentu, Dutch and English, and visitors add Spanish: the Central Bureau of Statistics credited 2025 growth to arrivals from South America, Canada and the United States. Have a native Papiamentu speaker write the first alert, then translate it. Test it by asking ten people to say back what they should do, and rewrite any line that fewer than nine get right.

What does a hotel or farm need before using AI to forecast water demand?

It needs a record, because a model cannot forecast from nothing. Start a spreadsheet with one meter reading a day, one rain-gauge reading and the number of guests or workers on site. After a full dry season you will have enough to see a pattern, and a utility, a university or a company can then test a simple forecast against it.