A Hybrid AI–Ontology Framework for Tropical Cyclone Forecasting, Risk Assessment, and Intelligent Disaster Decision Support

Tropical cyclones are among the most powerful forces on Earth, spinning up over warm ocean waters to unleash wind, rain, and storm surges that can reshape coastlines and threaten millions of lives. For decades, meteorologists have worked to predict where these storms will go and how strong they will become. These forecasts rely on complex computer models that track the storm's movement and intensity, giving authorities a vital head start. However, knowing a storm's path is only half the battle. A forecast that tells a city to expect high winds is useful, but it does not automatically tell officials which neighborhoods are most vulnerable, how many people need to leave, or what specific dangers—like flooding versus structural damage—await them. The gap between a raw weather prediction and a clear, actionable plan for saving lives has long been a challenge for disaster managers.

To bridge this gap, a team of researchers has developed a new system that combines the predictive power of artificial intelligence with a structured way of understanding the world, known as an ontology. Think of an ontology as a digital map of knowledge that connects facts about the storm to facts about the land and the people living there. While artificial intelligence is excellent at spotting patterns in massive amounts of data to predict the future, it often acts like a "black box," offering a number or a path without explaining the reasoning behind it. The researchers wanted to build a system that not only predicts where a cyclone will go but also translates that prediction into a clear, logical story about risk. By teaching a computer to understand the relationships between wind speed, rainfall, population density, and building strength, they created a tool that can generate specific, understandable advice for decision-makers.

The researchers tested this hybrid framework using three real-world cyclones that struck the Bay of Bengal: Amphan in 2020, Yaas in 2021, and Remal in 2024. The system operates in two main stages. First, it uses a type of artificial intelligence called a Long Short-Term Memory network to analyze historical weather data. This model looks at a sequence of past observations—such as the storm's location, speed, and pressure—to forecast its path for the next twenty-four hours. Simultaneously, a second artificial intelligence model estimates the size of the storm's impact zone, determining how far out from the center the dangerous conditions will reach. These predictions are not just numbers; they are fed into the second stage of the system, the ontology-based reasoning engine.

This reasoning engine acts as the translator. It takes the predicted path and size of the storm and layers them over a detailed digital profile of the affected region. This profile includes static information that does not change, such as the location of towns, the density of the population, the type of buildings in the area, and the local geography. The system then applies a set of logical rules to this combined data. For instance, if the AI predicts high winds and the ontology knows that a specific area has many old, fragile structures, the system infers a high risk of structural damage. If the forecast includes heavy rain and the area is low-lying, it infers a high flood risk. The result is a comprehensive risk assessment that goes beyond simple weather data to answer critical questions: How severe is the threat? Which areas are in the red zone? How many people are at risk?

When the team applied this framework to Cyclone Amphan, the most destructive of the three, the system performed with high accuracy. The AI predicted the storm's trajectory with an average error of about 72 kilometers over a twenty-four-hour period, a level of precision comparable to other advanced deep learning methods. More importantly, the reasoning engine correctly identified the event as a "Severe" threat, assigning it a red risk zone and recommending a high-alert response. It estimated that approximately 640,000 people were at risk, a figure that aligned closely with the massive emergency response that was actually launched during the real event. The system successfully distinguished between the extreme danger of Amphan and the slightly lower, though still significant, risks posed by Cyclones Yaas and Remal. For Yaas, the system also flagged a red zone and high alert, while for Remal, it correctly assessed a moderate risk, suggesting a yellow zone and a lower level of urgency.

The study demonstrates that combining predictive artificial intelligence with structured knowledge reasoning creates a more useful tool for disaster management than either approach could provide alone. The artificial intelligence handles the difficult math of predicting the storm's future, while the ontology layer ensures that these predictions are interpreted through the lens of local reality. This allows the system to produce outputs that are not just accurate numbers, but intelligible advice, such as specific evacuation directives or shelter advisories. The researchers found that this approach successfully transformed raw meteorological data into structured, semantic risk information that decision-makers could trust and act upon. While the current system relies on historical data and pre-defined rules, the authors suggest that future versions could incorporate real-time data from satellites and weather stations to make the system even more responsive. Ultimately, this work offers a promising path toward disaster management systems that are not only smart enough to predict the storm but wise enough to understand its true impact on the people and places it threatens.