An AI model could anticipate the development of flash floods: how it works

A review of more than 300 studies highlighted the use of algorithms in hydrological risk management and early warnings (Illustrative image Infobae) Artificial intelligence is no longer a tool limited to technology 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. It is also beginning to be applied to problems related to climate, where anticipating rapid changes can be decisive. In that field, a team from Penn State University developed a model designed to estimate, 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 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-Diffusion, a diffusion model, an artificial intelligence technique that learns to remove 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 transferred that principle to 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 following hours. The time scale is central to the problem. Flash floods can cause very high flow peaks in a few hours, even if the daily average does not seem exceptional. "Water levels can rise and fall rapidly within hours, which means that, at a daily scale, they may seem high, but not catastrophic," said Chaopeng Shen, lead author and professor of civil and environmental engineering, in a Penn State statement. To train h-Diffusion, the group used hourly flow records from 516 river basins in the United States collected between 1990 and 2003. Then, they tested the system with data from a later period, between 2009 and 2014, which the model had not used during its learning. The tool adapts the diffusion technique used in image generation to analyze rainfall and flow records (AP Photo) The paper indicates that the tool outperformed other advanced deep learning models used as benchmarks. Its advantage was not limited to flow estimation: it could also integrate recent measurements from monitoring stations without requiring complete retraining. That capability, called data assimilation, allows the system to incorporate information observed shortly before issuing a forecast. According to the university, the technique reduces some of the uncertainty associated with precipitation, whose hourly changes are difficult to anticipate. The model can use, for example, recent data on rainfall and flow to estimate what the day's water peak will be. In addition, the authors showed that the system can perform the inverse operation and estimate the probable hourly precipitation from the measured flow. "Ultimately, our final goal is to offer a robust, reliable system that can help save lives," said Shen, also affiliated with Penn State's Institute of Energy and the Environment, according to the statement. The team plans to analyze hybrid models and incorporate more physical variables into its predictions. Biases in the data and the demand 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 leverage artificial intelligence to improve flood risk management. A review published in Climate Risk Management, which examined more than 300 studies, 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 of areas exposed to flooding and help organize resources during an emergency. However, the review warns that its implementation faces obstacles. Among them are biases in the data, the lack of transparency of some models, and the demands of computing capacity. The authors proposed that the tools must be accessible and adaptable to different contexts, particularly in communities with fewer resources or limited monitoring systems. Physics-based weather models proved more accurate than algorithms in predicting unprecedented climate records (REUTERS/Kathleen Flynn) Another study, published in Science Advances, introduced a relevant warning for the use of these systems in weather alerts. The research compared AI models with HRES, a forecasting model based on physical principles from the European Centre for Medium-Range Weather Forecasts. The authors evaluated the systems' ability to anticipate records of heat, cold, and wind—that is, events that exceeded the values observed during the training period of the AI tools. In those cases, HRES performed better than the artificial intelligence models analyzed in most forecast time frames. 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 surpassed, 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.