Real-time Analytics
The bureau needed to replace slow, subjective manual observations of cloud shape, cloud amount, frost and other weather phenomena with a more continuous and accurate observation process.
Reported impact
The bureau created a practical AI weather-observation workflow that raised recognition accuracy high enough for operational use, reduced dependence on manual observers and opened a path to fully automated ground meteorological observation.
Hangzhou Meteorological Bureau built an AI observation workflow around field imaging and Baidu's EasyDL stack. High-definition fisheye cameras at the top of observation poles collect all-sky images, while ground-facing cameras capture grassland and surface conditions at the climate station. Those image streams are sent through APIs to custom EasyDL models trained on more than a million source images and a curated labeled set covering cloud forms, cloud amount, weather conditions, frost, rime and tea-frost scenarios. The bureau used PaddlePaddle-based hardware and Baidu's no-code AI development platform to package the recognition pipeline so it can run in real time and feed results directly into the meteorological information system. Instead of relying on observers to interpret conditions by eye, the bureau now has a repeatable image-driven process that supports forecast operations and moves weather observation toward full automation.
All-sky and ground images from 28 weather stations are captured by field cameras, sent by API to EasyDL recognition models running on the PaddlePaddle environment, and the returned labels for cloud, frost and weather conditions are written into the meteorological information system for forecasting use.
High-definition all-sky fisheye cameras and ground cameras collect images on a schedule, then call Baidu EasyDL model APIs for intelligent recognition (original: 图像数据通过API方式调用，实时调取百度大脑EasyDL模型接口进行智能识别).
Model recognition accuracy generally exceeded 85%, and cloud-form and cloud-amount recognition also passed 80% (original: 模型识别准确率普遍超过85%以上，其中包含20多个分类的云状识别及12个分类的云量识别准确率也在80%以上).
The bureau used image data from 28 stations, selected more than 11,000 images from over 1 million originals, and trained models for cloud, weather, frost and tea-frost recognition (original: 运用高清视频摄像机在28个气象站采集大量全天空、草面、树林、茶叶等图像数据，从100多万张原始图像中精选了11000多张).
Traditional manual observation could not meet the modern need for high-precision meteorological services (original: 传统人工气象观测已经无法满足高节奏的现代社会对高精度气象服务的需求).
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Baidu AI Cloud Sample Oct 12, 2025 Guiyang Public Security Traffic Management Bureau (贵阳市公安交通管理局) Guiyang needed a more systematic way to reduce congestion in a mountainous city where road geometry, rapid growth, and fragmented traffic-management methods made it hard to optimize signal timing and respond to changing… Average vehicle delay at target intersections fell 15%. AI Traffic Signal Control Large Model · Qianfan Application Development Platform · Qianfan Model Development Platform Baidu AI Cloud Sample Jun 19, 2025 Haikou Traffic Police Detachment After Typhoon Yagi damaged most field sensing equipment, Haikou traffic police urgently needed to restore signal-control effectiveness without waiting for every radar, camera, and detector to be rebuilt, or the city… 567 intersections were affected by the typhoon. Edge Computing Node · Qianfan AppBuilder · Qianfan ModelBuilder Baidu AI Cloud Sample Jun 19, 2025 Nanwei Software Co., Ltd. Nanwei needed a better way to help Fengning government users explore and analyze operational data because dashboards and cockpits could only display fixed views, and custom cross-team analysis cycles were too slow for… Baidu Jiuzhou Government Native Products · Qianfan AppBuilder · Qianfan ModelBuilder 