Flash floods, due to their sudden and localized nature, have long been the "ghost" disasters in global weather forecasting. Today, Google announced it has cracked this challenge through an innovative approach: using large language models to mine unstructured data from news reports, successfully building a global flood prediction system.

Traditional deep learning models often struggle to function in remote areas due to a lack of historical meteorological data. The Google research team changed their approach, leveraging the reading comprehension capabilities of to thoroughly analyze over 5 million news articles worldwide.

  • Data Conversion: The model extracted 2.6 million records of flood events from the news and transformed these qualitative descriptions into quantitative data with geographic tags and timestamps, forming a unique dataset called "Groundsource."

  • Model Training: Based on this "ground truth," researchers trained an LSTM neural network model that can predict the probability of flash floods in specific areas using global weather forecast data.

The head of Google's disaster resilience project stated that the greatest significance of the Groundsource dataset lies in its "balance."

  • Serving Vulnerable Areas: For countries and regions that cannot afford expensive weather radar systems or lack complete meteorological records, the model offers a low-cost early warning solution.

  • Field Validation: Currently, Google has already labeled risk levels for urban areas in 150 countries. Officials from the Southern African Development Community confirmed that the model has significantly improved local response speed to floods.

Although the model still has room for improvement in resolution (20 km) and real-time radar capabilities, this method of building a quantitative dataset from qualitative textual information opens up a new paradigm for disaster prevention and mitigation. The Google team said they plan to expand this technology to other short-lived but deadly natural disasters, such as heatwaves and mudslides.

By transforming AI's language understanding capabilities into early warning capabilities in the physical world, Google not only demonstrates the technical limits of , but also contributes a more inclusive technological force to global disaster prevention and mitigation efforts.