多源全天空图像及云掩膜标注数据集(2024-2025)OA
A dataset of multi-source all-sky images and cloud-mask annotations(2024-2025)
全天空成像是云量估计、短临辐射变化监测等研究与业务的重要数据来源.本全天空图像及云掩膜数据集包含三部分:其一,开源标注子集,对开源全天空图像数据集 Total-Sky Cloud Image Set(TCIS)进行统一的云/非云二值标注,共195 张;其二,开源全天空图像序列标注子集,基于 Atmospheric Radiation Measurement(ARM)体系的 Total Sky Imager(TSI)全天空图像序列,首先利用Segment Anything Model 2(SAM2)生成初始云掩膜结果,在此基础上由标注人员进行校正与复核,形成半自动云掩膜标注流程,进一步扩展得到连续时间序列的像素级标签,共 2560张;其三,站点观测序列子集,面向局地环境测量需求、支撑气象灾害等环境突发事件的快速响应,研究团队研制"地基多参数智能环境健康遥感诊断仪(地基康遥仪)".结合现有观测条件,为尽可能覆盖不同地理背景下的观测场景,选取分别位于高原湖区和高海拔高原的岱海与拉萨开展站点观测,获取 5 min 间隔连续全天空图像序列,分别为 1356张与 218张.数据生产实施多层次质量控制,面向云掩膜二值分割、跨数据源评测与深度学习提供数据支撑.
All-sky imaging provides hemispheric an important data source for cloud fraction estimation,short-term radiative variability monitoring and related research and operational applications.We release a dataset of all-sky images and cloud mask annotations comprising three components.(1)Open-source annotated subset:unified binary cloud/non-cloud annotations were generated for the open-source Total-Sky Cloud Image Set(TCIS),including 195 image-mask pairs.(2)Open-source all-sky image sequence annotation subset:We annotate all-sky image sequences acquired by the Total Sky Imager(TSI)within the Atmospheric Radiation Measurement(ARM)program using a semi-automatic pipeline.Initial masks are generated by the Segment Anything Model 2(SAM2)and subsequently corrected and reviewed by annotators,yielding 2,560 pixel-level masks for continuous time-series imagery.(3)Site-observation sequence subset:To meet local environmental measurement needs and support rapid responses to meteorological disasters and other environmental emergencies,we developed the Ground-Based Intelligent Multi-Parameter Environmental Health Remote Sensing Diagnostic Instrument(GB-EHRSDI).Given the available observational conditions,the Daihai and Lhasa sites,located in a plateau-lake region and a high-altitude plateau area,respectively,were selected to cover different geographical settings as broadly as possible.Continuous all-sky image sequences were acquired at a 5-minute interval,comprising 1,356 and 218 images,respectively.Multi-level quality control procedures were implemented during data production to support binary cloud-mask segmentation,cross-source evaluation,and deep learning.
张宇;徐磊;梁伟;徐敏;吴仕虹;王少华;曹春香
中国科学院空天信息创新研究院,遥感与数字地球全国重点实验室,北京 100101||中国科学院大学,北京 100094北京摩尔威视科技有限公司,北京 100084中国科学院大学,北京 100094||中国科学院空天信息创新研究院,微波成像技术全国重点实验室,北京 100101中国科学院空天信息创新研究院,遥感与数字地球全国重点实验室,北京 100101||中国科学院大学,北京 100094中国科学院空天信息创新研究院,遥感与数字地球全国重点实验室,北京 100101||中国科学院大学,北京 100094中国科学院空天信息创新研究院,遥感与数字地球全国重点实验室,北京 100101||中国科学院大学,北京 100094中国科学院空天信息创新研究院,遥感与数字地球全国重点实验室,北京 100101||中国科学院大学,北京 100094
全天空图像云掩膜像素级标注二值分割5 min连续观测
All-sky imagesCloud maskPixel-level annotationBinary segmentation5-minute continuous observations
《中国科学数据(中英文网络版)》 2026 (3)
334-343,10
国家科技重大专项(2025ZD01900407)国家重点研发计划(2023YFF0805904)国家自然科学基金(42471495).The National Science and Technology Major Project(No.2025ZD01900407)the National Key Research and Development Program of China(No.2023YFF0805904)the National Natural Science Foundation of China(Nos.42471495).
评论