{
  "id": 12473136,
  "title": "AI techniques behind image creation improve flash flood forecasts across U.S. river basins",
  "url": "https://urgent.news/2026/10/06/ai-techniques-behind-image-creation-improve-flash-flood-forecasts",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-06T21:30:01.000Z",
  "source": {
    "name": "Phys.org",
    "slug": "phys-org",
    "url": "https://phys.org/news/2026-10-ai-techniques-image-creation-river.html"
  },
  "original_language": "en",
  "account": "Flash floods pose a significant threat, capable of devastating communities in mere minutes. These catastrophic events account for nearly 85% of floods worldwide, causing thousands of deaths annually. Predicting flash floods is notoriously challenging, but a novel approach using artificial intelligence techniques, typically employed in image generation, may offer a solution.\n\nScientists at Penn State University, led by Professor Chaopeng Shen, have applied AI to improve hourly flash flood risk predictions. The team trained their model using rainfall and streamflow data collected from over 500 river basins across the continental United States between 1990 and 2003. Their innovative method demonstrates that the model can accurately parse unreliable weather data to generate precise predictions, and it can be easily updated with new data from water monitoring gauges.\n\nTraditionally, flood forecasting models utilize historical rainfall and streamflow data to make predictions. However, these models struggle to capture the severity of rapid weather changes that lead to flash floods, which often occur on an hourly scale. Shen and his team recognized this limitation and aimed to develop a more precise model by incorporating diffusion, a training method commonly used in image generation.\n\nIn their approach, the model is trained on a complete dataset or picture and then subjected to noise, which is subsequently removed to reconstruct an interpretable dataset. This technique allows the model to assimilate recent gauge observations without the need for retraining, a process known as inpainting. By incorporating real-time information, the model can provide more accurate predictions of flash flood risk.\n\nThe researchers trained their model on hourly streamflow records from 516 river basins across the U.S. from 1990 to 2003, subsequently testing its performance on data collected between 2009 and 2014. Shen and his team found that their diffusion approach generally outperformed existing hourly deep-learning models, particularly in predicting water flow during high-flow periods, which are notoriously difficult to forecast. Additionally, the model can run in reverse, taking recorded streamflow data at a river gauge and estimating the probable hourly rainfall that generated it.\n\nThe ultimate goal of this research is not to create the \"best\" model but to understand this technology and use it to help people. The researchers plan to continue developing their model, exploring hybrid approaches to training AI models and incorporating more physical variables in predictions. Ultimately, they aim to create a robust, reliable system that can save lives by providing accurate flash flood forecasts.",
  "summary": "A flash flood can devastate a community for years in a matter of minutes. These disasters account for roughly 85% of floods and cause thousands of deaths annually worldwide, with water levels capable of shooting up 30 feet (9 meters) and receding to normal levels in the span of a single, unpredictable day, according to the National Weather Service.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}