基于多重网格策略的NLS-3DVar资料融合方法及其在气温数据融合中的应用

NLS-3DVar Data Fusion Method Based on Multigrid Implementation Strategy and Its Application in Temperature Data Fusion

将多重网格策略引入NLS-3DVar(Non-linear Least Squares-based on Three-dimensional Variational Data Assimilation,非线性最小二乘三维变分同化)方法,进而应用于2400多个国家级气象观测站逐时气温数据和NCEP再分析气温数据的融合,得到中国区域空间分辨率1°×1°,时间分辨率为6小时的气温融合产品。分别从单重网格(分辨率1°×1°)和双重网格(分辨率由2°×2°到1°×1°)利用2014年1~12月(4、5月除外)的独立检验数据考察NLS-3DVar气温融合产品质量,验证基于多重网格策略的NLS-3DVar方法的优越性。在单重网格下,与广泛应用于气象行业的Cressman插值产品(均方根误差和相关系数的年平均值分别为1.961°C d^(-1)和0.924)相比,NLS-3DVar产品全年始终具有最小的均方根误差和最大的相关系数,年平均值分别为1.915°C d^(-1)和0.929;站点间误差分析进一步表明,NLS-3DVar产品在大多数检验站点精度更高,在新疆、甘肃、云南、陕西等地区尤为突出;加入双重网格策略的NLS-3DVar产品与单重网格的NLS-3DVar产品误差对比显示,均方根误差年平均值分别为1.649°C d^(-1)和1.711°C d^(-1),相关系数年平均值分别为0.970和0.968,二者在均方根误差和相关系数的表现上都极为相似,即双重网格NLS-3DVar气温产品尽管对观测数据采取了稀疏化处理,但依旧维持了原有的产品精度,并且在计算效率上提高了1倍多。而与同样在双重网格下基于多尺度的STMAS(Space–Time Multiscale Analysis System)算法相比,双重网格的NLS-3DVar产品在产品精度上同样占据优势,在计算效率上单位时次耗时与STMAS算法几乎相当。

In this study, the authors first incorporate the multigrid implementation strategy into the non-linear least squares-based three-dimensional variational data assimilation system （NLS-3DVar）, which is applied for temperature data fusion. A merged temperature dataset at 1° resolution and 6-hour interval is produced based on in situ observations at 2400 observational sites over China and NCEP （National Centers for Environmental Prediction） final global tropospheric analyses. Another set of independent validation data （from January to December except April and May in 2014） is used to evaluate the merged dataset. The dataset of NLS-3DVar is compared with the gridded data at 1° resolution produced by the widely used Cressman interpolation method. NLS-3DVar product always has lower RMSE （Root Mean Square Errors） of 1.961℃ d-I and higher correlation coefficient of 0.924 compared to the dataset produced by Cressman interpolation. The precision of merged temperature product of NLS-3DVar is higher in most stations and independent of validation data, especially at those stations in Xinjiang, Gansu, Yunnan, Shanxi, and so on. The performances of NLS-3DVar based on both the single grid and multigrid strategies are also compared. Both RMSE and correlation coefficient have little differences. Although multigrid-based NLS-3DVar uses the sparse process, the precision is almost the same as that of single-grid based NLS-3DVar. However, its computational costs are greatly reduced due to the sparse process. Compared with the STMAS algorithm （Space-Time Multiscale Analysis System）, multigrid-based NLS-3DVar performs better regardingthe precision of product with almost the same computational efficiency.