静止卫星地表温度数据是研究昼夜气候和环境变化的重要参数。但现有发布的静止卫星地表温度数据由于受到云等大气因素的影响，往往出现数值缺失现象。针对该问题，提出基于昼夜变化模型的风云静止卫星地表温度空值数据的稳健修复方法。由多项式、傅里叶函数和高斯函数构建新的昼夜变化模型，并利用Levenberg-Marquardt算法进行模型参数的求解与优化，进而实现空值修复。以风云2号F星数据（FY-2F）为例，模拟不同类型的像元缺失情况进行修复，并将不同模型修复结果与真实温度值比较，同时也对真实数据进行了测试。结果表明：本文提出的修复方法能有效对温度空值数据修复，且优于传统方法。

Land Surface Temperature (LST) derived from geostationary satellites (GEO-LST) is one of the key parameters in analyzing diurnal climate and environment changes. Compared with polar-orbiting satellite data, the LST of geostationary satellites has become increasingly attractive because of its high temporal resolution. GEO-LST has been widely applied in meteorological, climatological, and hydrological studies. However, the existing GEO-LST products can suffer from cloud cover, cloud contamination, and atmospheric disturbance, resulting in missing data. The objective of this study is to develop a new method for reconstructing Feng Yun geostationary satellite (FY-2F) LSTs based on Diurnal Temperature Cycle (DTC) and robust regression.
The hourly LSTs from FY-2F with 5 km spatial resolution are adopted for this study, and different types of missing pixels are simulated and reconstructed based on a new model. This model is composed of polynomial, Fourier, and Gaussian functions (PFG), and a robust regression called Levenberg-Marquardt (LM) algorithm is adopted for solving and optimizing the modelparameters. A performance test is conducted by comparing the PFG model based on the LM algorithm with the other two methods implemented on FY-2F data, namely, a method based on Least Square (LS) algorithm modeling of the previous model (van den Bergh. F VAN2006) and a method based on LS algorithm modeling of the PFG. Moreover, the new method is also tested for real LST products (with missing data).
By comparing the reconstructed results of different methods (LM-PFG, LS-PFG, and LS-VAN2006) with the real LST values, the simulated experiment indicates that the LM-PFG method obtains the lowest root mean square compared with the other two methods. The stability analysis of these methods shows that LM-PFG is the least sensitive to missing samples, which validates that LM-PFG is more stable than LS-PFG and LS-VAN2006. Simultaneously, experimental results show that LM-PFG can also reconstruct the missing values with high accuracy and better than the other methods.
In this study, an approach for reconstructing GEO-LSTs based on DTC and robust regression is proposed. The method is evaluated using simulated and real FY-2F LST products and can obtain high scores on quantitative evaluation and visual connectivity. Further work can be conducted to expand this method to other potential GEO-LSTs, such as Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager LSTs or Geostationary Operational Environmental Satellite Imager LSTs. Similar to previous studies, the reconstructed LST values of the proposed method are estimated under cloud-free conditions. LSTs under cloudy conditions should be investigated in future studies.