In a talk at the 54th Annual Conference on Science and the Environment, Gila Loika of Google Research presented artificial intelligence systems for warning of earthquakes, fires, hurricanes and floods. While impressive progress has been made, the systems do not predict every disaster with certainty and do not replace emergency authorities.
“Our goal is that no one will be surprised by a natural disaster – at least not by a disaster that is approaching,” she said. Gila Loika, Product Manager for Flood Forecasting and Crisis Resilience inGoogle ResearchIn her talk at the 54th Annual Conference on Science and the Environment, she presented Google's attempt to turn smartphones, satellites, and models of artificial intelligence For a global warning system.
The conference was held on July 8 and 9, 2026, at the Congress Center in the United Nations Buildings in Jerusalem. The event, described as the largest environmental forum in Israel, was led by the Israel Society for Ecology and Environmental Sciences, hosted by the Jerusalem Municipality and in collaboration with the Hebrew University. ([Hayadan – Hayadan][1])
Loika began by referring to the deadly earthquakes that had recently struck Venezuela. She said that such events illustrate how science and technology can provide the public with critical seconds or hours, even when the disaster itself cannot be prevented. The article is based on the transcript and summary of the lecture.
Phones are becoming a network of earthquake sensors
One example she presented is Android's earthquake alert system. Accelerometers installed in smartphones and typically used to detect the device's orientation and count steps can also detect movement characteristic of an earthquake.
When a large number of devices in a given area simultaneously detect a similar pattern of tremors, the system analyzes the signals and estimates the location of the event. It can then warn users further away from the epicenter before the strong seismic waves reach them.
Loika emphasized that this is not about predicting an earthquake before it starts. The system detects a tremor that has already occurred and takes advantage of the fact that electronic information travels faster than seismic waves. Sometimes it is only a few seconds, but they may be enough to move away from windows, stop machinery or take cover. ([blog.google][2])
The Carmel fire led to a focus on fires
According to Loika, the field of fires is particularly close to Google's Israeli team. The team in Haifa watched the deadly Carmel fire unfold in 2010, and the event influenced their decision to delve deeper into developing tools for mapping crises and disseminating information during emergencies.
Google already displays approximate boundaries of fires in search and on maps. However, the accuracy of the mapping and the speed of updates depend on the satellites passing over the area. Some existing observation systems provide an up-to-date image only a few times a day – a very long period of time when a fire front is moving quickly.
"When firefighters protect our homes, communities and families, the greatest gift we can give them is time," said Loika.
For this purpose, a system is established FireSat – A dedicated satellite array for detecting and monitoring fires. The day before her talk, on July 7, the first three operational FireSat satellites were launched, after an experimental satellite was launched in March 2025. They were built by Muon Space and are operated as part of the Earth Fire Alliance, with support from entities including Google.org.
When the full array, planned to include about 50 satellites, is operational, the goal is to identify fire hotspots as small as five by five meters and provide new global observations every 20 minutes. The three satellites launched now are only the first phase, so this does not mean that such updates are already being received everywhere in the world. ([Earth Fire Alliance][3])
Dozens of scenarios instead of a single forecast
Another area is WeatherNext, a family of models developed by Google DeepMind and Google Research for weather forecasting. Instead of producing just one forecast, the models can quickly calculate a collection of possible scenarios. This allows us to estimate not only the most likely scenario, but also the likelihood of rare and devastating events.
Loika gave the example of Hurricane Melissa, which made landfall in Jamaica in October 2025 as a Category 5 hurricane. When the storm was still classified as a weak tropical depression, conventional models projected different scenarios. Five days in advance, WeatherNext estimated an 80% chance that the system would strengthen and hit Jamaica as a Category 5; three days before impact, the probability rose to nearly 100%.
The forecast did not replace human forecasters. It provided the U.S. National Hurricane Center with an additional signal, helping it issue more decisive warnings and giving authorities and residents time to evacuate and protect their property. ([Google DeepMind][4])
Seven days before a river floods, a day before a flash flood
Google operates inFlood Hub River flood forecasts up to seven days in advance. According to the company, the forecasts cover areas at significant risk, home to approximately two billion people in more than 150 countries.
But river flooding is only part of the problem. Flash floods They develop rapidly, sometimes hitting urban areas shortly after heavy rain. For years, it was difficult to train models for them because there was no uniform global database documenting where and when such floods occurred.
To address the lack of data, researchers developed the GroundsourceGemini analyzed news reports collected over a period of about 20 years and converted them into a structured database containing about 2.6 million records of flood events. The database was used to train a model that calculates the probability of flash flooding in an urban area over the next 24 hours. ([research.google][5])
The model also has limitations. At this point, it operates at a resolution of about 20 by 20 kilometers and focuses mainly on urban areas. More news reports are published about cities than about rural and sparsely populated places, so the training dataset is not equally representative of all parts of the world. Even a real event that was not reported in the media may appear to the model to have been wrongly predicted when evaluated. ([research.google][6])
The time to act before disaster strikes
Loika mentioned a collaboration with GiveDirectly in Nigeria, where early forecasts were used to transfer money to families before flooding hit. The money allowed them to evacuate, buy sandbags, or move equipment and belongings to safety. This is an example of “preemptive action” – moving from waiting for a disaster and distributing aid afterward, to taking action when the forecast indicates real danger. ([sites.research.google][7])
The systems she presented do not eliminate uncertainty. They do not prevent a hurricane, fire or earthquake, and may miss events or issue warnings that turn out to be unnecessary. Their role is to reduce the information gap and give the public and emergency services a resource that cannot be produced after the disaster has already begun – time.
"Technology can't always prevent natural hazards," concluded Loika, "but through better science, strong partnerships, and responsible artificial intelligence, we can help ensure that people get the information they need before danger arrives."
Questions and Answers
Can Google predict earthquakes? No. The Android system detects an earthquake after it has started and sends an alert to areas where the seismic waves have not yet reached.
What is FireSat? A satellite array designed to detect and map fires using multispectral sensors and artificial intelligence. The full array is expected to provide global observations every 20 minutes.
How far in advance can floods be predicted? Flood Hub displays forecasts for river flooding up to seven days in advance and for urban flash floods up to 24 hours in advance.
How was the flash flood database created? Groundsource used Gemini to extract structured information from news reports collected over about 20 years, creating approximately 2.6 million event records.
More on the subject on the science website
- Flood prediction using machine learning with over 90% accuracy
- Google DeepMind reveals: A fifty-year leap forward in weather forecasting
- Mobile phones can warn of forest fires and extreme weather conditions
- Adapted models for predicting forest fires significantly improve warning capabilities
- NASA: Climate crisis doubles the extremes of forest fires worldwide
For the original publication: Opening the scientific article
