Why India's Most Advanced Weather Models Still Send Alerts Late

The recent floods in Kerala have exposed a worrying duality in India’s disaster management response. 

On one hand, the state government insists it has learned its lessons and is now planning tech-enabled solutions, with Chief Minister VD Satheesan announcing an AI-backed disaster management programe with advanced forecasting and flood mapping. On the other, current systems—the base on which these advanced tech will be built—fail to deliver even basic alerts.

SACHET, the National Disaster Alert Portal, is designed to flash geo-targeted emergency alerts to cell devices alerting of oncoming severe weather. However, a fisherman along Kerala's coast told a TV channel the alert reached his phone only after rough weather had already set in. On other days, his phone lights up with warnings meant for faraway districts, while his own stretch of coastline stays quiet until the weather turns.

Those concerns have echoed through fishing communities since Cyclone Ockhi in 2017. Nearly a decade later, the same questions resurfaced during Kerala's flash floods this August, when parts of the state received over 300 mm of rainfall on a single day. The Indian Meteorological Department (IMD) maintained it had issued Orange Alerts well ahead of the event. However, residents still felt caught off guard. CPI(M)'s Pathanamthitta district secretary, Raju Abraham, alleged that no warning preceded the release of water from the Moozhiyar Dam, contributing to flash floods that left over 3,000 people stranded.

The question now is whether an accurate forecast can survive institutions, technology, and communication networks before disaster strikes, and not after.

That distinction matters because India has spent the past decade investing heavily in forecasting technology. They include higher-resolution numerical weather prediction models, Doppler radars, dense observation networks, satellite monitoring, and machine-learning pilots. Platforms like NDMA's Sachet alert system, the IIT Madras-developed Bharat Forecast System, and the Ministry of Earth Sciences' Mission Mausam all point toward an increasingly AI-enabled weather ecosystem. 

Yet Kerala's floods suggest better models alone don't produce better warnings.

From Accuracy to Usefulness

S Abhilash, a scientist at the Advanced Centre for Atmospheric Radar Research, Cochin University of Science and Technology, believes accuracy is the wrong metric to measure.

"We should speak less about forecast accuracy and more about forecast usefulness," he tells AIM.

The distinction is subtle but significant. Forecast accuracy determines whether scientists correctly predicted rainfall, wind speed, or storms. Forecast usefulness, meanwhile, asks whether that information reached the people who needed it, in a form they could understand, and early enough to plan ahead. The two don't always align.

As part of an ongoing research programme, Abhilash and his colleagues work directly with fishing communities along Kerala's Thiruvananthapuram coast, including Vizhinjam, Marina, and Karingalum, and with farming communities in Wayanad. Rather than generating new forecasts, they take information already issued by agencies like IMD and the Indian National Centre for Ocean Information Services, combine multiple datasets, and translate it into advisories tailored to how each community actually makes decisions.

A fisherman deciding whether to head into the Arabian Sea needs very different information from a farmer planning irrigation. However, they often get the same district-wide bulletin.

"The right forecast has to reach the last mile at the right time," Abhilash says. "Generating a good forecast serves no purpose if it doesn't reach the people who actually need it."

That shift from one-size-fits-all forecasts to stakeholder-specific warnings is also where AI could play one of its most useful roles. Rather than just improving prediction itself, researchers increasingly see AI's value in integrating multiple data streams, personalising warnings, and translating technical forecasts into advice specific users can act on. Instead of a blanket rainfall warning for a district, an AI-assisted system could eventually distinguish between a fisherman heading offshore, a farmer in a low-lying paddy field, and a family living beside a river, sending each a different advisory calibrated to their risk.

Where the Warning Chain Breaks

Abhilash frames forecasting not as a single technological achievement but as an end-to-end system with three stages.

The first is observation: satellites, Doppler radars, automated weather stations, ocean buoys, and river gauges, increasingly supplemented by new sensors and machine-learning techniques capable of processing huge volumes of data.

The second is prediction—numerical weather models converting those observations into forecasts ranging from hours to more than a week out.

The third is communication, and this, Abhilash believes, is where even an excellent forecast starts to lose its value. Technical terms get oversimplified, bulletins get delayed, some people receive repeated alerts for weather that doesn’t affect their locality, while others hear nothing until the danger is at the door.

This leads to alert fatigue and citizens start losing trust in the system.

But there a slow, if certain, shift from weather forecasting to impact forecasting.

A traditional forecast answers a scientific question like ‘how much rain will fall’. Then the IMD issues alert on the basis of rainfall in a particular district. A Red, Yellow, or Orange alert is defined by expected rainfall thresholds—but whether that rainfall causes minor disruption or catastrophic flooding depends on the vulnerability of the landscape.

Impact-based forecasting asks a different one: what will that rainfall do? The outcome varies from one topography to the next, depending on terrain, drainage, river levels, soil saturation, and how built-up the area is. 

Can AI Fix the Last Mile?

If forecasting is only one part of the warning chain, where does AI actually fit?

The country now has many of the building blocks of an AI-enabled early warning system: NDMA's Common Alerting Protocol-based Sachet platform for geo-tagged alerts, dense observations through the Weather Information Network Data System (WINDS), river-gauge and dam-release data, multilingual dissemination channels, and next-generation forecasting systems like the Bharat Forecast System.

"The technical capability is largely there," remarks Prof A Damodaran of IIM Bangalore. "The real bottleneck is making sure the information reaches every household in a form that people can act upon."

In his view, AI shouldn't be seen simply as a forecasting tool but as a way to integrate information across agencies, identify vulnerable communities, personalise alerts, and help emergency managers decide faster. Yet, even though it may warn a family about the floodwater, someone still has to help them understand when to leave, where to go, and what to take.

"Human agency remains the last layer," Prof Damodaran says. 

He believes the next frontier for AI is micro-level vulnerability data. Rainfall alone tells only part of the story. Two houses on the same street can see very different damage depending on elevation, construction, drainage, and surrounding land use. Feeding that kind of hyperlocal detail into AI systems could produce far more credible impact forecasts, and eventually enable parametric insurance that pays out based on expected damage rather than lengthy claims assessments. 

The System is Still Catching Up

On the forecasting side, India has made real technical progress. 

The Bharat Forecast System, developed at IITM Pune and run on the Arka supercomputer, operates at roughly 6 km resolution—twice as fine as the previous Global Forecast System—letting forecasters simulate atmospheric conditions in far greater detail. 

But a higher-resolution model doesn't automatically produce a better warning on the ground. A forecast still has to move through a long institutional chain: numerical models produce rainfall predictions that meteorologists interpret, river-flow models estimate flood risk, agencies decide whether to issue warnings, state disaster authorities prepare a response, telecom operators and broadcasters disseminate the alert, and finally, a resident decides whether to trust the message enough to act. 

Every step is a place for delay, distortion, or loss. 

For Neetha K. Gopal, Director of IMD's Kerala Meteorological Centre, much of the public confusion comes down to a misunderstanding the operational processs.

"The preparation of Quantitative Precipitation Forecasts (QPF) and Flash Flood Guidance issued by IMD based on Numerical Weather Prediction models remains a fully manual operational process," she says.

The Ministry of Earth Sciences is running pilot projects to improve accuracy using AI, but those remain experimental rather than part of routine operations.

Gopal also draws a clear line between weather forecasting and flood warnings. "Flood forecasting is done by the Central Water Commission (CWC) using Quantitative Precipitation Forecasts provided by IMD, real-time hydro-meteorological observations and river runoff models," she explains.

On the delay between forecasting floods and the public alert reaching residents, she says, "Flood warning is the responsibility of CWC." But Sachet National Disaster Alert Portal remains the only portal to publish official warnings.

That institutional split is a fuse box waiting to short during a natural disaster. There is no single agency responsible for the entire journey from forecast to public action. 

Among the most affected by an unsyncronised response are the fishing communities. Several fishermen say alerts arrive late, or are issued for districts other than the waters they're actually working.

Gopal rejects the suggestion that warnings weren't disseminated. According to her, IMD began issuing wind warnings covering the Kerala and Karnataka coasts and Lakshadweep from July 28, updated four times daily, and distributed through All India Radio, Doordarshan, the Department of Fisheries, KSDMA, and other print and broadcast media.

That leaves an open question: if the forecast existed and the advisory was issued, where did the communication chain snap? Or did a district-wide advisory simply fail to reflect the hyperlocal conditions? 

The Road to Hyperlocal Forecasts

Forecasting agencies are recognising the gap. IMD is working toward more hyperlocal warnings by expanding observation networks and computing capacity under Mission Mausam to strengthen weather and climate research. Combined with AI-assisted forecasting and denser observations, such efforts could eventually produce warnings specific not just to a district but to a stretch of coastline, a river basin, or a single vulnerable neighbourhood.

Researchers like Abhilash argue that's precisely where forecasting needs to go—not simply more accurate, but more meaningful.