摘要:基于格点统计插值分析系统（Gridpoint Statistical Interpolation analysis system，简称GSI），利用粒子滤波（Particle Filter，简称PF）方法对卫星红外辐射率资料进行了云覆盖、云高等三维云图产品的反演研究。选取了具有高时空分辨率的静止卫星GOES（Geostationary Operational Environmental Satellites）-Imager辐射率资料进行了云反演试验，初步评估了PF云反演方法的可行性及其与多元极小残差（Multivariate and Minimum Residual，简称MMR）云反演方法的异同。结果表明：两种方法反演得到的云覆盖和云顶气压与NASA基于CO2切片法反演得到的GOES云产品一致性较高。PF和MMR方法反演产品的优点是云图信息是三维分布的，相对于NASA提供的GOES云产品能提供更全方位立体的云信息。MMR方法需要利用一维变分逐步拟合观测来反演三维云图产品；PF方法采用不同模式垂直层的云覆盖比例作为不同粒子来近似后验概率分布，计算效率大大提高。进一步提出了一种新的基于“扰动粒子”的粒子滤波云反演方法，结果表明：在滤波过程中采用足够多的粒子样本（样本数量约为250）可以改进后验概率密度函数的估计，有效地避免了粒子发散问题，改善了云反演的结果。
关键词:地球静止卫星成像仪（GOES-Imager）/
格点统计插值分析系统（GSI）/
云反演方法/
粒子滤波法
Abstract:The Particle Filter (PF) cloud retrieval method developed in the framework of GSI (Gridpoint Statistical Interpolation analysis system) is able to directly utilize the Infrared radiances to retrieve cloud masks and cloud profiles. Cloud retrieval experiments with GOES (Geostationary Operational Environmental Satellites)-Imager radiance are conducted with the PF and the Multivariate and Minimum Residual (MMR) methods respectively for comparison. The retrieved cloud properties from both methods show a good agreement with the cloud products from GOES. MMR retrieves cloud fractions on each individual model vertical levels by minimizing a cost function, while PF is an effective algorithm to treat those cloud fractions as different particles to gain recursive estimations of cloud distributions. To improve the PF method in terms of cloud retrieval application, this study perturbs the particles to better estimate cloud distributions. The advanced PF (with roughly 250 samples) is appropriate to ameliorate the problem of filter divergence caused by limited particles with better cloud retrievals efficiently.
Key words:Geostationary Operational Environmental Satellites (GOES)-Imager/
Gridpoint Statistical Interpolation system (GSI)/
Cloud retrieval methods/
Particle filter
PDF全文下载地址:
http://www.iapjournals.ac.cn/dqkx/article/exportPdf?id=93cfc148-973f-4727-9722-2e3912c16a39
