基于梯度提升树的大气分层透过率快速计算方法

A Gradient Boosting Tree method for rapid calculation of level-to-space transmittances

中文关键词:  大气透过率  红外辐射传输  机器学习  梯度提升树

英文关键词:Atmospheric transmittance  Infrared radiative transfer  Machine learning  Gradient Boosting Tree

中文摘要:

      大气透过率的计算是红外辐射传输计算的核心，RTTOV（Radiative Transfer for TOVS）通过建立大气廓线中温度、水汽、臭氧和其他气体浓度等参数与卫星通道透过率的统计关系，可实现卫星通道透过率和大气顶辐射率的快速准确计算。但在一些复杂吸收波段，如水汽波段，RTTOV的计算误差较大。为提高RTTOV在水汽敏感波段的计算精度，利用机器学习中的梯度提升树（Gradient Boosting Tree，GBT）方法，选取从ECMWF（European Centre for Medium-Range Weather Forecasts）的IFS-137（The Integrated Forecast System，137-level-profile）廓线集中挑选的1406条廓线和由此计算的透过率真值作为样本，选取风云三号气象卫星上搭载的红外分光计（InfraRed Atmospheric Sounder，IRAS）通道12（7.33 μm）进行个例研究，分别建立陆地和海洋晴空大气等压面至大气层顶透过率的快速计算模型（GBT模型）。通过和透过率、亮温真值的比较，验证了GBT模型。比较结果显示，GBT模型预测的透过率平均绝对误差（Mean Absolute Error，MAE）为：陆地0.0012，海洋0.0009；均方对数误差（Mean Squared Logarithmic Error，MSLE）为：陆地0.0215，海洋0.0095，均小于RTTOV直接计算的透过率的误差（陆地、海洋的MAE分别比RTTOV小0.0008和0.0010，MSLE分别比RTTOV小0.0135和0.0227）；由GBT模型计算的亮温MAE分别为：陆地0.0949 K，海洋0.0634 K，均方根误差（Root Mean Square Error，RMSE）分别为：陆地0.1352 K，海洋0.0831 K，也都小于RTTOV直接模拟的晴空亮温误差（陆地、海洋的MAE分别比RTTOV小0.1685 K和0.1466 K，RMSE分别比RTTOV小0.1794 K和0.1685 K）。本研究的结果表明，在IRAS红外水汽波段，GBT预测的透过率和亮温误差比RTTOV小。机器学习有提高水汽波段正演精度的潜力，或可为辐射传输的快速计算提供可行的替代方法。

英文摘要:

      The calculation of atmospheric transmittance is the core of solving emitted infrared radiative transfer equations. As the layer transmittance of the atmosphere is parameterized by functions of the mean layer temperature，water vapor，ozone and other gas concentrations，RTTOV (Radiative Transfer for TOVS) is able to quickly and accurately compute level-to-space transmittance and top-of-atmosphere radiance. However，large computing errors have been found in some strong absorption bands，for example，the water vapor band. To solve this problem，a machine learning method called Gradient Boosting Tree (GBT) is utilized to compute transmittances in this paper. 1406 typical profiles have been selected from the ECMWF IFS-137 (European Centre for Medium-Range Weather Forecasts，the Integrated Forecast System，137-level-profile) as training samples. The water vapor channel (7.33 μm) of the IRAS (InfraRed Atmospheric Sounder)/FY-3 (Fengyun 3 series) is selected for a case study. A fast model to compute clear-sky level-to-space transmittance and brightness temperature by the GBT method (GBT model hereafter) has been built. The calculated transmittances and brightness temperatures have been validated by the ground-truth. The comparison results show that the mean absolute errors (MAE) of clear-sky transmittances calculated by the GBT model are 0.0012 (land) and 0.0009 (ocean)，while the mean squared logarithmic errors (MSLE) are 0.0215 and 0.0095 over land and ocean，respectively. These results are smaller than those calculated by RTTOV (MAEs of 0.0008 (land) and 0.0010 (ocean)，MSLEs of 0.0135 (land) and 0.0227 (ocean)). In addition，the MAEs of clear-sky BT calculated by the GBT model are 0.0949 K (land) and 0.0634 K (ocean)，and the root mean square errors (RMSE) are 0.1352 K (land) and 0.0831 K (ocean)，respectively，which are smaller than those by RTTOV (MAEs of 0.1685 K (land) and 0.1466 K (ocean)，RMSEs of 0.1794 K (land) and 0.1685 K (ocean)). The case study in this paper demonstrates that in the water vapor band of IRAS，the transmittance and brightness temperature predicted by the GBT are more accurate than those simulated by RTTOV. Machine learning method has the potential to improve the accuracy of the transmittance and brightness temperature in the water vapor band. It provides an optional solution to the fast radiation transfer calculation.