基于快速局域线性回归的IRAS/FY-3B大气温湿廓线反演

提出一种快速的局域线性回归(Fast Locally Linear Regression,FLLR)算法,用于从搭载在风云三号B星(FY-3B)上的红外大气探测仪(IRAS)红外观测数据反演大气温湿廓线。算法所需的观测样本为IRAS/FY-3BL1数据红外观测值与AIRX2RET V5产品的时空匹配数据,以2011年为例,在180°W~180°E、60°N~60°S的研究区域内按照观测时间绝对差小于15min和观测角度绝对差小于2°的条件获取观测样本,并对样本进行了评价。在匹配观测样本的基础上比较分析了FLLR算法与LLR算法、D矩阵算法和非线性的神经网络算法,然后采用FLLR算法从IRAS/FY-3BL1数据反演得到2011年全年的大气温湿廓线,并外推反演得到2012年第1季度的大气温湿廓线。最后,利用相应的ECMWF再分析数据和RAOB探空观测对2011年的反演结果进行了精度验证,采用AIRX2RET V5产品对2012年第1季度的外推反演结果进行了验证。结果显示:与D矩阵算法相比,FLLR算法反演大气温度和湿度廓线的均方根误差分别减小~0.8K和~0.5g/kg,其精度与非线性的神经网络算法相当;相对于ECMWF再分析数据,本文大气温度和湿度廓线反演结果的均方根误差分别小于2.5K和2.3g/kg;而相对于RAOB数据,其均方根误差分别小于3.5K和2.0g/kg;2012年第一季度外推反演结果的均方根误差分别小于2.5K和1.6g/kg,与算法精度基本一致。IRAS/FY-3B大气温湿廓线的反演精度与MOD07V5大气廓线产品相当。

Inversion of IRAS/FY-3B atmospheric temperature and humidity profiles based on fast locally linear regression

The FLLR algorithm was developed to retrieve atmospheric temperature and humidity profiles from IRAS/FY-3B data.The observation samples required by the algorithm are IRAS /FY-3B L1 observation value and time-spacing matching data of AIRX2RET V5. Take the year of 2011 for an example, the observation sample was obtained in the survey region of 180°W~180°E and 60°N~60°S under the condition that absolute difference of the observation time is less than 15min and absolute difference of the observation angle is lower than 2°, and then an evaluation on the sample was conducted. Based on the observation sample matching, a comparison and analysis was carried out on FLLR algorithm and LLR algorithm, D Matrix algorithm and non-linear neural network algorithm. Then the year-round atmospheric temperature and humidity profiles in 2011 were obtained from IRAS/FY-3B L1 data inversion by using the FLLR algorithm, and the profile in the first quarter of 2012 can be achieved through extrapolation inversion. The last step was a precision test on the inversion results of 2011 by means of corresponding ECMWF reanalysis data and RAOB radiosonde observation, as well as verification on the extrapolation inversion results in the first quarter of 2012. It demonstrates that compared with D Matrix algorithm, the root mean square error in the inversion of the atmospheric temperature and humidity profile by using the FLLR algorithm has respectively reduced by ~0.8K and ~0.5g/kg, while its precision is quite the same as the non-lineared neural network algorithm. With respect to ECMWF reanalysis data, the root mean square error of the atmospheric temperature and humidity profile inversion is respectively lower than 2.5 K and 2.3 g/kg, and for RAOB data, the root mean square error is lower than 3.5 K and 2.0 g/kg respectively. The root mean square error of the extrapolation inversion in the first quarter of 2012 is respectively lower than 2.5 K and 1.6 g/kg, and the precision is almost the same in different algorithms. The inversion precision of IRAS/FY-3B atmospheric temperature and humidity profile is quite the same as the precision of MOD07 V5 atmospheric profile.