The artificial intelligence It is no longer a tool limited to technological laboratories. Its systems intervene in everyday tasks, organize information, automate processes and expand capabilities in areas such as medicine, industry, education, transportation and scientific research. Also begins to be applied in problems related to the climatewhere anticipating rapid changes can be decisive.
In that field, a team of the Pennsylvania State University developed a model aimed at estimating, hour by hour, the risk of flash floods. The proposal adapts an AI method used to create images and seeks to more accurately recognize sudden increases in river flow, which can appear and disappear in a short time.
Scientists at Penn State University developed an artificial intelligence model to forecast flash floods (AP Photo)
The work was detailed in a study published in the journal Water resources research. There, the researchers presented h-diffusiona diffusion model, a artificial intelligence technique that learns to eliminate random information to reconstruct recognizable patterns from complex data.
In image generation, this system can start from an image with noise, that is, with altered visual information, and recover a coherent version. The team carried that principle into flood forecasting: Instead of reconstructing a photograph, the model analyzes rainfall and flow records to estimate how a river’s water will vary over the next few hours.
The time scale is central to the problem. Flash floods can cause very high flow peaks in a few hours, although the daily average does not seem exceptional. “Water levels can rise and fall rapidly in a matter of hours, meaning that, On a daily scale, they may seem high, but not catastrophic“said Chaopeng Shen, lead author and professor of civil and environmental engineering in a statement from State of Pennsylvania.
to train h-diffusionthe group used hourly flow records 516 river basins in the United States collected between 1990 and 2003. He then tested the system with data from a later period, between 2009 and 2014, that the model had not used during its training.
The tool adapts the diffusion technique used in generating images to analyze rainfall and flow records (AP Photo)
The paper indicates that the tool outperformed other advanced deep learning models used as a reference. Its advantage was not limited to flow estimation: it was also able to integrate recent measurements from monitoring stations without requiring complete new training. That capability, called data assimilation, allows the system to incorporate information observed shortly before issuing a forecast.
According to the study house, The technique reduces part of the uncertainty associated with precipitationwhose time changes are difficult to anticipate. The model can use, for example, recent rainfall and flow data to estimate what the peak water will be that day. Furthermore, the authors showed that the system can perform a reverse operation and estimate the probable hourly precipitation from the measured flow.
“Ultimately, our ultimate goal is to offer a robust, reliable system that can help save lives,” said Shen, also affiliated with the Energy and Environment Institute of State of Pennsylvaniaaccording to the statement. The team plans to analyze hybrid models and incorporate more physical variables into their predictions.
Biases in data and the requirement for computing capacity represent the main obstacles to the implementation of this technology (REUTERS/Athit Perawongmetha)
The advance is part of a field that seeks to take advantage of the artificial intelligence to improve flood risk management. A review published in Climate risk managementwho examined more than 300 investigations, identified applications in risk assessment, early warning systems and emergency response.
That analysis argues that algorithms can process large volumes of information, from hydrological records and weather forecasts to satellite images and sensor data. The combination of these sources can provide faster estimates on flood-prone areas and help organize resources during an emergency.
However, the review warns that its implementation faces obstacles. These include data biases, lack of transparency in some models, and computing power requirements. The authors stated that tools must be accessible and adaptable to different contextsparticularly in communities with fewer resources or with limited monitoring systems.
Physics-based weather models demonstrated greater accuracy than algorithms in predicting unprecedented climate records (REUTERS/Kathleen Flynn)
Another study, published in Scientific advancesintroduced a relevant warning for the use of these systems in weather alerts. The investigation compared AI models with HRESa forecast model based on physical principles from the European Center for Medium-Range Weather Forecasts.
The authors evaluated the systems’ ability to anticipate record heat, cold and wind, that is, events that exceeded the values observed during the training period of the AI tools. In those cases, HRES outperformed the AI models analyzed across most forecast timelines.
The work concludes that AI models tended to underestimate both the intensity and frequency of unprecedented extreme events. The greater the margin by which a record was broken, the greater the bias of the automated predictions.
The difference may be linked to the way both approaches work. Artificial intelligence models learn from patterns present in historical data, while physics-based systems use equations that describe the evolution of the atmosphere. Therefore, the authors point out that Rigorous evaluations are still needed before using AI tools exclusively in critical early warning and disaster management systems.
