防洪韧性的下一篇章：开源谷歌水文框架
内容总结：
谷歌开源洪水预测模型，助力全球防灾减灾
2026年6月3日，谷歌研究院宣布将其自主研发的水文模型框架正式开源，旨在帮助各国气象水文部门将先进的人工智能洪水预报技术整合到自身工作流程中。这一举措被视为全球洪灾防御能力的一次重要升级。
洪水是全球最具破坏性的自然灾害之一，往往预警时间短、破坏力持久。多年来，谷歌研究院持续打造基于人工智能的高精度洪水预报模型，并在谷歌洪水预警平台（Flood Hub）投入使用。此次开源的水文模型框架，让研究人员和预报员能够利用与Flood Hub相同的架构和训练数据，自主训练AI洪水预报模型。
该开源模型是一个Python软件包，基于PyTorch机器学习框架开发。它通过获取气候、土壤、地形、植被等地理特征数据，结合降雨、气温等气象预报信息，预测全球河流的每日流量。模型架构采用长短期记忆网络，可利用开源数据集Caravan的历史河流数据进行训练。各国专家还可根据本地流域特点，添加专属数据对模型进行微调，实现“因地制宜”的精准预警。
值得注意的是，此次开源发布包含两种版本：一是2024年基准研究中测试的原始模型，二是目前支撑Flood Hub实时全球洪水预报的升级版。最新的基准研究表明，升级版模型在已有观测数据流域将可靠预测时间窗口延长了6天，在无观测数据流域也延长了1天。
谷歌与捷克水文气象研究所的深度合作，充分验证了这一开源工具的实际应用潜力。捷克水文气象研究所不仅确认AI模型预报质量已达到传统物理模型水平，还专门开发了适配器，将开源模型接入全球广泛使用的洪水预报平台Delft-FEWS。这意味着各国水文部门无需更换现有系统，即可轻松集成AI技术。
世界气象组织水文模拟与预报部门负责人金辉麟博士对此表示欢迎，称赞开源水文建模工具“对于支持社会水资源管理和应对环境挑战至关重要”，并强调这将有助于实现“让每个社区都能提前获得灾害预警”的全球使命。
目前，该模型架构、完整文档及培训资源已以Apache 2.0开源许可协议形式发布在GitHub上。谷歌表示，希望通过这一举措，让资源有限的地区和团队也能借助先进AI技术进行高效洪水预报，无需昂贵的基础设施投入，真正实现防灾减灾技术的普及与民主化。
中文翻译：
2026年6月3日
格雷·尼尔林与黛博拉·科恩，谷歌研究部研究科学家
我们已开源水文模型，使国家气象水文部门能够将先进的AI洪水预报技术整合到自身工作流程中。
洪水是全球最具破坏性的自然灾害之一，往往预警时间极短却造成长期损害。多年来，谷歌研究部构建了先进的AI模型以实现更精准的洪水预报，确保这项技术能触达一线应急人员，为他们争取行动时间。为帮助进一步保护脆弱社区，我们现已在GitHub上开源水文建模框架，供他人使用和改进。
这一开源建模框架使研究人员和预报员能够基于与谷歌洪水中心河流洪水预报相同的架构和相似训练数据，训练AI洪水预报模型。其设计初衷是让水文学者能够在我们谷歌研究部的工作基础上，通过添加和测试新模型、新数据及新方法进行拓展。该框架还允许作业预报员——即负责为特定区域提供可操作的洪水预警的专业人员——将本地数据和知识融入基于AI的先进洪水预报中。
我们相信，当科研成果能够赋能他人复现并拓展研究发现，确保创新成为全球进步的催化剂时，才算真正发挥其全部潜力。正因如此，我们在内部开发了该框架，并与捷克水文气象研究所（CHMI）等合作伙伴进行了测试。开放模型架构与训练流程，代表了全球洪水防御能力的根本性转变：既能让国家气象水文部门及其他气象机构、主管机关完全掌控自有数据，又能赋能本地专家利用专业化数据集优化模型。
我们的水文模型是一个Python软件包，采用开源PyTorch机器学习建模包，实现了驱动谷歌洪水中心的河流预报模型。这些模型输入与气候、土壤、地形和土地覆被相关的地理特征数据，结合降雨、气温及其他天气状况的气象预报数据，从而预测全球河流的日流量。
该水文建模软件包包含基于长短时记忆网络（LSTM）的模型架构，以及可利用开源Caravan数据集中的历史河流数据进行模型训练的训练流程。研究人员和洪水预报机构可向这一开源数据仓库添加自有数据，针对本地流域训练或微调模型。
如需开始实施，请查阅Python交互式教程笔记本，以及相关视频教程（YouTube平台），了解如何操作模型代码。
该代码仓库包含水文模型的两个不同版本：一版为2024年发布的基准测试研究中验证的原始版本，另一版为当前驱动洪水中心实时全球洪水预报的升级模型。新模型基于初期研究的奠基性成果，过渡至新的模型架构。该框架使我们能够将多样化的多源气象输入整合为统一的洪水预测系统（如下图所示）。我们近期的基准测试研究表明，与前版相比，新模型在有观测流域将可靠预测窗口延长了6天，在无观测流域延长了1天。
在《2025年多灾种预警系统全球现状》报告中，世界气象组织指出，本地数据及本土与地方知识是有效灾害预警的关键组成部分，并强调“将本土与地方知识系统性地纳入风险知识生产仍是例外而非常态”。我们的开源洪水预报工作流程直面报告发现，使区域预报员能够直接、亲手掌控AI驱动的预报模型。这些框架训练相对简便且成本低廉，既具备传统水文预报模型无法比拟的精准度，又无需其复杂性，同时允许用户纳入自有专业数据进行训练和预测。
易采用的开源工具对于弥合技术创新与洪水灾害系统实际效能之间的鸿沟至关重要，尤其在加速预警系统能力建设方面。
此次开源的运营潜力，最佳例证莫过于我们与CHMI的合作。双方协作的关键在于验证了我们的AI模型能够提供与本地校准的传统概念模型质量相当的预报。CHMI还开发了一个适配器，将水文开源框架集成至Delft-FEWS平台——这是国家和地方洪水预报机构、非政府组织及私营企业广泛使用的主流运营洪水预报工具，用于驱动预测模型。Delft-FEWS由三角洲研究院运营维护。这一集成使CHMI及全球其他水文机构能在标准工作流程中运用该模型。它亦为全球机构将机器学习纳入水资源管理工作流程提供了蓝图。
除CHMI这类大型机构外，开源模型发布还提供了可规模化、易获取的工具，让先进预报技术惠及大众，为资源受限地区和本地团队打开了大门，使其能够利用高质洞察力，无需昂贵传统预报基础设施。
国际气象学界已认可这一开放科学方法的价值。世界气象组织水文建模与预报科科长金辉麟博士指出：“开源水文建模工具的扩展对于支持社会管理水资源和应对环境挑战至关重要，我对此表示欢迎。世界气象组织致力于支持开源、可互操作、由成员国驱动的模型和工具，这些工具能够拯救生命，并推动确保全球社区提前收到灾害预警以保护生命和生计的全球使命。”
模型架构、完整文档及培训材料现已在GitHub上以Apache 2.0许可协议发布，使研究人员和运营预报专业人员均可完全访问该框架。
通过将水文模型交予全球水文学界之手，我们能够共同构建一个更具洪水韧性的世界。关于谷歌更广泛洪水预报计划与资源的更多详情，请访问谷歌研究部网站。我们诚邀全球水文学界在此基础上进一步开发这些开源工具。
众多人员参与了此项工作的开发。我们特别感谢CHMI的雅库布·克雷伊奇和扬·丹海尔卡在合作与反馈中的贡献，以及来自谷歌研究部和社会影响合作团队的以下同仁：阿米特·马克尔、阿维纳坦·哈西迪姆、黛博拉·科恩、艾米莉·莱因斯坦、吉拉·洛伊克、格雷·尼尔林、妮娜·贝克莱、奥姆里·谢菲、鲁文·萨亚格、罗尼·阿米拉、什穆利克·弗罗曼、斯蒂芬妮·里斯和约西·马蒂亚斯。
英文来源：
June 3, 2026
Grey Nearing and Deborah Cohen, Research Scientists, Google Research
We have open-sourced our hydrology model to enable National Meteorological and Hydrological Services to integrate advanced AI-based flood forecasting into their own workflows.
Floods are one of the most devastating natural hazards worldwide, often arriving with little warning and leaving long-term damage. Over several years, Google Research has built state-of-the-art AI models for more accurate flood forecasting, ensuring this technology reaches frontline responders to give them time to act. To help further protect vulnerable communities, we are now open-sourcing our hydrology modeling framework on GitHub for others to use and build upon.
This open source modeling framework allows researchers and forecasters to train AI flood forecasting models with the same architecture and similar training data to what is used to power riverine flood forecasts on Google’s Flood Hub. It is developed to allow hydrological scientists to build on what we have done at Google Research by adding and testing new models, data, and approaches. It also allows operational forecasters — people whose job entails providing actionable flood warnings for specific areas — to incorporate local data and knowledge into state-of-the-art AI-based flood forecasting.
We believe that a scientific breakthrough reaches its full potential when it empowers others to replicate and expand upon findings, ensuring that innovation is a catalyst for worldwide progress. That's why we developed this framework internally and tested it with partners like the Czech Hydrometeorological Institute (CHMI). Releasing our model architecture and training pipeline represents a fundamental shift in global flood preparedness, allowing National Meteorological and Hydrological Services (NMHSs), other meteorological agencies, and authorities to retain full control of their data while empowering local experts to refine models using specialized datasets.
Our hydrology model is a Python package that uses the open source PyTorch machine learning modeling package to implement the river forecast model that drives the Google Flood Hub. These models take input data in the form of geographical features related to climate, soils, topography, and land cover, along with meteorological forecasts related to rainfall, temperature, and other weather conditions to predict the daily flow rate of rivers around the world.
The hydrology modeling package includes model architectures based on Long Short Term Memory (LSTM) Networks, and a training pipeline that allows these models to be trained using historical river data from the open source Caravan dataset. Researchers and flood forecasting agencies can add their own data to this open source data repository to train or fine tune models to their local watersheds.
To get started with implementation, check out this interactive tutorial notebook in Python and the associated video tutorial on Youtube on navigating the model code.
This code repository includes two distinct versions of our hydrological model: the original version tested in our benchmarking study published in 2024, and an upgraded model that currently powers real-time global flood forecasting on Flood Hub. The new model builds upon the foundational success of our initial research by transitioning to a new model architecture. This framework allows us to process diverse, multi-source meteorological inputs into a unified flood prediction system, illustrated in the figure below. Our recent benchmarking study shows that this new model extends the reliable predictive horizon by six days in gauged basins and by one day in ungauged basins relative to the previous version.
In the Global Status of Multi-Hazard Early Warning Systems 2025 report, the World Meteorological Organization recognizes that both local data and Indigenous and Local Knowledge (ILK) are critical components of effective disaster warnings, and notes that “[t]he systematic integration of ILK into risk knowledge production is still the exception rather than the norm.” Our open source flood forecasting workflow addresses the report’s finding by allowing regional forecasters to take direct, hands-on control over AI-powered forecasting models. These frameworks are relatively easy and inexpensive to train, providing accuracy without the complexity of traditional hydrological forecasting models and allowing users to incorporate their own specialized data for training and prediction.
Readily adoptable open-source tools are critical for bridging the gap between technological innovation and the real-world effectiveness of flood hazard systems, particularly for accelerating capacity development around early warning systems.
The operational potential of this release is best illustrated by our partnership with CHMI. Their collaboration was key to validating that our AI-based model provides forecasts comparable in quality to traditional, locally calibrated conceptual models. CHMI also developed an adapter that integrates the hydrology open source framework into the Delft-FEWS platform, a popular operational flood forecasting tool used by national and local flood forecasting agencies, NGOs, and private companies to drive predictive models. Delft-FEWS is operated and maintained by the Deltares research institute. This allows CHMI and other hydrological services worldwide to use the model in their standard workflows. This integration serves as a blueprint for how global agencies can include machine learning in their water management workflows.
Beyond larger institutions like CHMI, the open source model release offers a scalable, accessible tool, democratizing access to advanced forecasting and opening the door for resource-constrained regions and local teams to leverage high caliber insights without the need for costly traditional forecasting infrastructure.
The international meteorological community has recognized the value of this open-science approach. Dr. Hwirin Kim, Chief of Hydrological Modelling and Forecasting Section at the World Meteorological Organization, notes: “I welcome the expansion of open-source hydrological modeling tools that are critical to supporting how societies manage water resources and respond to environmental challenges. We at WMO are keen to support open-source, interoperable, Member-driven models and tools that can help save lives and advance the global mission to ensure communities everywhere are forewarned about hazards to protect their lives and livelihoods.”
The model architecture, comprehensive documentation, and training materials are now live on GitHub under an Apache 2.0 license, making the framework fully accessible to both researchers and operational forecasting professionals.
By putting our hydrology model into the hands of the global hydrology community, we can build a more flood-resilient world. More details about Google’s broader flood forecasting initiatives and resources are available on the Google Research site. We invite the worldwide hydrology community to build upon these open tools.
Many people were involved in the development of this effort. We would especially like to thank Jakub Krejci and Jan Daňhelka from CHMI for their partnership and feedback as well as the following individuals across Google Research and the Social Impact Partnerships Team: Amit Markel, Avinatan Hassidim, Deborah Cohen, Emily Reinstein, Gila Loike, Grey Nearing, Nina Bekele, Omri Shefi, Reuven Sayag, Rony Amira, Shmulik Fronman, Stephanie Rees, and Yossi Matias.