Paper Title: AI expands high-quality urban flash flood forecasts globally
Authors: Oleg Zlydenko et al.
Affiliated Institution: Google Research / Jigsaw
Publication Timing: 2026 (preprint)
Research Field: AI + Climate Disaster Prediction + Earth Science + Data-Based Disaster Management

 As the frequency and intensity of natural disasters increase due to climate change, the importance of disaster prediction technology is also rapidly growing. The research "AI expands high-quality urban flash flood forecasts globally," published in this context, is a study that proposes a system capable of predicting flash floods in urban areas worldwide using artificial intelligence. This paper, which involved Google Research and Jigsaw researchers, is evaluated as a case demonstrating that AI can develop from a simple data analysis tool into a disaster response infrastructure.

The core message of the paper is relatively clear. By utilizing AI technology, high-quality flood forecasting systems can be established even in areas with insufficient sensor infrastructure. This presents not just a technical achievement but a possibility to change the structure of disaster response systems.

There are three important contemporary changes in the background from which this research emerged. The first is the increase in natural disasters due to climate change. According to the research, approximately 85% of floods worldwide are flash floods and more than 5,000 people die from such disasters annually. Flash floods are extremely fast-occurring and locally manifesting, so response time is extremely short. Due to these characteristics, flash floods are evaluated as one of the most dangerous natural disasters.

The second background is global inequality in disaster warning systems. In developed countries, sophisticated early warning systems utilizing radar, hydrological sensors, and high-resolution weather models have been established. However, many developing countries and regions in Africa and South Asia lack such infrastructure. According to research, disaster mortality rates in countries lacking early warning systems appear to be approximately 6 times higher. That is, the problem lies not simply in the occurrence of natural disasters but in the inequality of technology infrastructure.

The third background is the emergence of AI-based earth system research. In recent years, artificial intelligence has been rapidly introduced into meteorology and earth science research. With representative AI-based weather prediction models like GraphCast and various climate AI models appearing, AI is establishing itself as an important tool for understanding earth systems. This paper also emerged as research within this trend.

Conventional flood forecasting methods had a relatively traditional structure. It was a method of collecting sensor data, inputting it into hydrological models, and then predicting flood occurrence probability. However, this approach has a fundamental limitation: sensors are not installed in many areas and data itself is often insufficient.

The research team proposed a new approach to solve this problem — specifically, learning using news article data. The researchers extracted flood events from news articles worldwide and used this as training data. A dataset built by extracting flood-related information from more than 5 million news articles was established under the name "Groundsource."

This dataset was created by automatically extracting flood occurrence events using large language models. This demonstrates that unstructured text data can be used as training data for disaster prediction models. While sensor data alone was previously the primary input data, now text data such as news can also be an important information source.

The core AI model used in the paper is LSTM (Long Short-Term Memory). LSTM is a deep learning model with strengths in analyzing time-series data and is frequently used to analyze time-based data such as weather data.

The model utilizes various data simultaneously. First, it uses static data such as terrain information, land characteristics, and population density. This data was built based on the AlphaEarth Foundation model. It also utilizes various weather data as input variables, including ECMWF weather models, NASA satellite precipitation data, and NOAA precipitation data. Here, weather records from the past 7 days are analyzed together to predict the probability of flooding occurring over the next 24 hours.

The research team compared the AI model's performance with the National Weather Service's flood warning system. The results were interesting. The AI model's prediction performance appeared at a level similar to existing specialized systems. This means that a considerable level of flood prediction is possible with AI alone, without expensive sensor networks or radar systems.

The technical characteristics of this research can be summarized into three main points. The first is the news-based dataset. The Groundsource dataset was built by extracting flood events from more than 5 million news articles, demonstrating that unstructured text data can be utilized in disaster prediction systems.

The second is the global learning model. This model does not learn from data from a specific country but simultaneously learns from worldwide data. This structure enables prediction even in regions with insufficient data.

The third is the low-cost system structure. Conventional flood prediction systems required large-scale infrastructure such as radar, sensor networks, and national meteorological agencies. In contrast, this AI system can operate with satellite data and global weather models alone.

The most important significance of this research is the democratization of disaster response technology. In the past, flood prediction technology was close to being the exclusive domain of developed countries. High-cost sensor networks, weather radar, and specialized personnel were required. However, using AI-based models, warning services can be provided to more than 150 countries with a single system. This demonstrates the possibility of significantly expanding access to disaster technology.

This research also presents a new role for AI. Recent AI discussions often focus on generative AI or automation technology. However, this research demonstrates that AI can be utilized as public infrastructure that protects human lives.

Another important change is the transformation of the data paradigm. In the past, sensor data was the core input data for models. However, this research demonstrates that text data such as news articles can also be utilized in earth system analysis. That is, it has presented the possibility that large language models and news data can be utilized as a new form of earth system data.

Of course, several limitations also exist in the research. The current spatial resolution of the model is approximately 20 km, which may limit very precise regional prediction. Also, news-based data may be affected by reporting bias or regional disparities. Additionally, the current dataset has the problem of not being able to clearly distinguish between urban flooding and river flooding.

Nevertheless, this research demonstrates an important change. Artificial intelligence is expanding into areas such as earth system management and disaster response beyond technology used simply for advertising recommendations or content analysis.

In the past, AI was mainly utilized in commercial areas such as recommendation algorithms or advertising systems. However, AI has now begun to be used for issues of human survival such as climate change response, disaster prediction, and environmental management.

The most important message this paper delivers is clear. AI can become not merely a smart technology but social infrastructure that reduces risk and protects lives.