迈向一个无人对自然灾害感到惊讶的世界
内容总结：
谷歌发布AI防灾新进展：目标“无人因自然灾害措手不及”
随着全球极端天气与自然灾害频发，谷歌团队近日在“AI为地球”活动上宣布，其利用人工智能在灾害预测、监测与应急响应领域取得突破性进展，致力于实现“无人因自然灾害措手不及”的愿景。
洪水预警覆盖150国，城市内涝可提前24小时预报
谷歌自2018年在印度巴特那试点以来，基于机器学习的全球洪水预测模型已取得重大突破。其“洪水中心”现覆盖超过150个国家、20亿人口，河流洪水可提前7天预报。针对城市突发内涝，新开发的AI模型“Groundsource”基于20年公开数据训练，可提供最长24小时的预警。相关数据集与水文框架已开源，供全球科研与应急机构使用。
野火监测扩展至34国，卫星星座可识别5米火点
谷歌利用卫星图像在搜索与地图中提供AI野火边界追踪，目前已覆盖34个国家，今年新增7国。其参与研发的“FireSat”卫星星座计划由50余颗卫星组成，可每20分钟更新一次，探测地球表面小至5米×5米的火点。首颗原型卫星已于去年入轨。
飓风与极端天气预测精度提升
新一代模型“WeatherNext 2”可在数分钟内生成全球逐小时高精度预报，预测风速、降水、气压等关键变量。在2025年飓风季中，该模型成功提前数天高置信度预测飓风路径与强度，协助牙买加气象部门提前5天向公众发出预警。
热浪与空气质量警报触达超百国
谷歌在搜索中推出极端高温警报，覆盖100多个国家并提供防护建议；安卓地震预警系统可在地震波到达前向用户发送警报；地图应用在30多国提供实时空气质量数据。
政企合作强化全球韧性
谷歌与联合国机构、多国政府及救援组织合作，将灾害信息通过搜索、地图及公共警报系统推送至数十亿用户。在尼日利亚和孟加拉国，救援机构利用谷歌洪水预报提前发放紧急现金，帮助民众撤离。谷歌旗下慈善机构Google.org持续资助各地灾后重建。
谷歌气候与地理空间AI模型已集合为“谷歌地球AI”数据集，支持商业与政府机构应对灾害响应与行星监测等挑战。谷歌表示，将继续推进AI研究，与全球伙伴携手，让自然灾害不再成为“意外”。
中文翻译：
迈向一个无人因自然灾害而措手不及的世界
全球极端天气事件和自然灾害急剧增加，给社区带来毁灭性打击。过去十年，谷歌团队致力于在危机时刻——往往是人们最需要的时候——提供有用的信息。
我们推动了基于人工智能的突破性研究，从提供及时信息发展到预测和探测野火、洪水、地震和极端天气等自然灾害。我们通过数十亿人使用的谷歌产品提供关键信息，并与全球政府及组织合作，帮助社区为这些危机做好准备并应对。
危机时刻的可操作信息有助于拯救生命和生计：我们抗灾能力的北极星是，没有人应该因自然灾害而措手不及。
在今天的“AI for the Planet”活动中，我们分享了如何朝着这一愿景迈进，将人工智能驱动的工具和洞察交到合作伙伴和用户手中。以下是我们取得的进展以及未来的展望。
推进预测与探测
十年前，大规模的可靠洪水预测在很大程度上被认为遥不可及。我们实现洪水预测全球影响力的多年历程始于2018年在印度巴特那地区的试点，其假设是借助机器学习，我们能够大规模预测洪水。自那时起，我们逐步推进研究并扩大部署。凭借发表在《自然》杂志上的河流洪水全球模型突破，我们将范围扩展至数据稀缺地区；通过基于人工智能的新方法Groundsource，我们利用20年的公开报告构建了高质量洪水数据集，并用于训练山洪模型。如今，洪水中心的预报覆盖150多个国家、20亿人口，覆盖面临重大洪水风险的区域。河流洪水预报可提前7天提供，而针对城市地区的新山洪预测能为这些突发事件提供长达24小时的提前预警。我们已开源山洪数据集和水文框架，以便研究人员、企业和当地专家构建新的解决方案。
对于气旋等极端天气事件，WeatherNext 2提供了迄今最准确的预测。它能在数分钟内生成全球范围内高度详细的逐小时预报，并预测风速、风向、降水和气压等关键天气变量。在2025年飓风季期间，它成功提前数天高置信度地预测了气旋的路径和强度。
对于野火，我们利用卫星图像在搜索和地图中提供基于人工智能的边界追踪。自早期工作以来，我们已将覆盖范围扩展至34个国家，今年新增7个。为提升未来火灾探测能力，我们与地球火灾联盟和Muon Space合作，在谷歌.org、摩尔基金会、贝佐斯地球基金等的资助下共同开发了FireSat。首颗原型卫星去年已进入轨道。由50多颗卫星组成的完整FireSat星座将能探测到地球上任何地方仅5x5米的野火，并每20分钟更新一次。
针对极端高温，我们正将人工智能应用于卫星和航空图像，绘制城市环境中建筑的反射率（正如我们刚发表的成果）。这有助于城市了解如何通过使用凉爽屋顶来降低地表温度。
关键时刻的实时警报与权威信息
我们在搜索和地图上通过SOS警报提供危机响应更新，整合来自官方和可信媒体的相关信息。我们与90多个国家的授权警报发起方和分发方合作，通过公共警报系统放大紧急警报和公共警告。我们的危机信息被浏览数十亿次；仅去年一年，谷歌平均每天帮助人们连接危机信息超过1000万次。
信息要有用，必须可操作。例如，搜索上的极端高温警报为100多个国家的人们提供警告，包括全球高温健康信息网络的安全提示。安卓地震警报系统在地震到达前检测地震并提醒安卓用户，为人们争取安全避险时间。谷歌地图在30多个国家提供最新空气质量数据，帮助用户减少污染物暴露。
对全球共同使命的持续支持
建设全球抗灾能力需要合作。通过与政府、联合国机构、组织、科学家和急救人员合作，我们可以帮助各地社区免受自然灾害侵害。
在尼日利亚和孟加拉国，GiveDirectly和国际救援委员会利用我们的洪水预测采取预防行动，在洪水上涨前发放应急现金，使社区能够疏散并保护财物。在飓风梅丽莎期间，美国国家飓风中心使用我们的WeatherNext模型，提前五天预测了其在牙买加登陆，使牙买加气象局能够通知公众。在全球范围内，谷歌.org正与当地组织合作并资助灾后恢复工作。
过去十年，我们在推动基于人工智能的气候韧性研究突破和解决方案方面取得了进展，为全球社区提供了可操作的及时信息。我乐观地认为，通过利用人工智能并与合作伙伴协作，我们将更接近一个无人因自然灾害而措手不及的世界。
英文来源：
Towards a world where no one is surprised by a natural disaster
The world is experiencing a dramatic rise in extreme weather events and natural disasters, devastating communities. Over the past decade, our teams at Google have worked to make helpful information available to people at times of crises — often when they need it most.
We’ve advanced AI-based breakthrough research and progressed from providing timely information to forecasting and detecting natural disasters such as wildfires, floods, earthquakes and extreme weather. We’ve made critical information accessible via Google products that are used by billions, and partnered with governments and organizations around the world to help communities prepare for and respond to these crises.
Actionable information in times of crises can help save lives and livelihoods: our north star for our crisis resilience efforts is that no one should be surprised by a natural disaster.
At today’s AI for the Planet event, we shared how we’re making progress towards this vision, putting AI-powered tools and insights in the hands of our partners and users. Here’s a look at how we got here and what’s ahead.
Advancing forecasting and detection
A decade ago, reliable flood prediction at scale was largely considered out of reach. Our multi-year journey to global impact in flood forecasting began with a pilot in the Patna region in India in 2018, and the hypothesis that with machine learning, we could help predict floods at scale. Since then, we’ve progressively advanced research and scaled deployment. With our global model breakthrough for river floods, published in Nature, we expanded to data-scarce regions, and with our new AI-based methodology, Groundsource, we built a high-quality floods dataset based on 20 years of public reports, which we used to train a flash floods model. Today, forecasts on Flood Hub cover 2 billion people across more than 150 countries, in areas at risk for significant flood events. River flood forecasts are available up to seven days in advance, and our new flash flood predictions in urban areas provide up to 24-hour advance notice of these rapid-onset events. We’ve open sourced both the flash floods dataset and our hydrology framework, so researchers, businesses and local experts can build new solutions.
For extreme weather events like cyclones, WeatherNext 2 delivers our most accurate predictions yet. It can generate highly detailed hourly forecasts for the whole globe in minutes, and is capable of forecasting crucial weather variables including wind speed and direction, precipitation and pressure. During the 2025 hurricane season, it successfully predicted the path and intensity of cyclones with high confidence days in advance.
For wildfires, we use satellite imagery to provide AI-based boundary tracking in Search and Maps. Since our early work, we have expanded to provide coverage in 34 countries, including seven new countries this year. To improve future fire detection capabilities, we co-developed FireSat in collaboration with the Earth Fire Alliance and Muon Space, supported by funding from Google.org, the Moore Foundation, the Bezos Earth Fund and others. The first protoflight satellite was placed in orbit last year. A full FireSat constellation of 50+ satellites would be able to detect wildfires just 5 x 5 meters anywhere on earth, with updates every 20 minutes.
To address extreme heat, we’re applying AI to satellite and aerial imagery to map the reflectivity of buildings across urban environments, as we just published. This can help cities understand how to reduce surface temperatures by using cool roofs.
While individual models are powerful, many real-world questions require a holistic approach. Answering complex queries like, "Where is a hurricane likely to make landfall, and which communities are most vulnerable and how should they prepare?" requires reasoning about imagery, population and the environment. We’ve brought together our climate and geospatial models in the Google Earth AI collection of models and datasets. It enables planetary intelligence and is helping businesses and organizations address challenges like disaster response and planetary monitoring.
Real-time alerts and authoritative information when it matters most
We provide crisis response updates on Search and Maps with SOS alerts, which bring together relevant information from authorities and trusted media outlets. And we partner with authorized alert originators and distributors in over 90 countries to amplify emergency alerts and public warnings with Public Alerts. Our crisis information has had billions of views; last year alone, Google helped connect people with crisis information over 10 million times per day, on average.
For information to be useful, it must be actionable. So for example, Extreme heat alerts on Search provide warnings for people in over 100 countries, including safety tips from the Global Heat Health Information Network. The Android Earthquake Alerts System detects earthquakes and alerts Android users before shaking reaches them, to give people time to get to a safe place. Up-to-date air quality data is available on Google Maps in over 30 countries, to help users reduce their exposure to pollution.
Ongoing support for a shared global mission
Building global resilience requires collaboration. By working with governments, UN agencies, organizations, scientists and first responders, we can help keep communities everywhere safe from natural disasters.
In Nigeria and Bangladesh, GiveDirectly and the International Rescue Committee have used our flood forecasts to power anticipatory action, distributing emergency cash ahead of rising waters so that communities can evacuate and safeguard their belongings. During Hurricane Melissa, when the U.S. National Hurricane Center used our WeatherNext model, it predicted the Jamaican landfall five days ahead, enabling the Met Service in Jamaica to notify the public. And across the world, Google.org is partnering with local organizations and funding disaster recovery efforts.
Over the past decade, we've made progress driving AI-based research breakthroughs and solutions for climate resilience, providing actionable, timely information to communities around the world. I'm optimistic that by harnessing AI and working with our partners, we’ll move closer towards a world where no one is surprised by a natural disaster.