1998-2012年青藏高原TRMM 3B43降水数据的校准

运用1998-2012年青藏高原的TRMM 3B43降水数据以及气象台站实测降水数据,对比分析了青藏高原地区TRMM 3B43降水数据偏差分布规律。结果表明:(1) TRMM 3B43降水数据在青藏高原地区存在明显误差,特别是降水量大的地区和月份,偏差量较大。(2)青藏高原地区TRMM 3B43降水数据偏差分布与海拔、经纬度、降水量存在密切的关系。用偏差分布规律,加法修正法结合随机森林算法对青藏高原地区TRMM 3B43降水数据进行了校准。经过校准之后,数据精度得到显著提高,有效增加了数据的可用性,多年月平均数据决定系数R²最大可达到0.9(3、10月),最小也接近于0.5(12月),效率系数E均为正值,最大可达到90(3、10月);多年季平均和多年平均降水数据中除了第一季度结果稍差外(决定系数R²为0.58),其余数据校准效果均较好。

Calibration of TRMM 3B43 over Tibetan Plateau during 1998-2012

Precipitation is a vital element of the water cycle in the Earth System, which is closely related to the ecological, hydrological, and meteorological processes. Because of the special topographical and climatic conditions, the Tibetan Plateau has great influence in regional and atmospheric circulation patterns. And the hydrology cycle in the Plateau is an important aspect of study for thermal forcing effect on regional climate. Furthermore, various surface features on the Plateau(e.g. vegetation)have a close relationship with precipitation. So it is meaningful to study the precipitation of the Tibetan Plateau. Owing to the lack of ground observations, it is difficult to study the spatial and temporal patterns of precipitation over the plateau. The satellite remote sensing technology can be utilized to fill in the gaps where station data are not available. There are already a series of regional and global remote sensing precipitation products. The Tropical Rainfall Measuring Mission(TRMM)is one of the highest resolution(0.25°×0.25°)products among all the current satellite precipitation datasets. Unfortunately, because of the influence of the terrain and atmospheric condition, the accuracy of it is not very ideal in northwest mountainous areas of China. So it is critical to find a method to calibrate it for further use. Comparing to rain-gauges observed precipitation dataset collected from China Meteorological Data Sharing Service System, the TRMM 3B43 precipitation data over the Tibetan Plateau have some deviation. This paper analyzed the deviation and found it closely related to altitude, latitude, and longitude and precipitation distribution. Supposing that the rain-gauges observed precipitation data are accurate, point-based deviation between original TRMM precipitation and rain-gauge observed precipitation can be calculated. Then surface-based deviation can be obtained using some interpolation algorithm such as Kriging. What is noted above is called addictive correction method. In this paper, originally TRMM 3B43 precipitation data are calibrated using the additive correction model and random forest algorithm. The Random Forest (RF) is constructed based on the classification and regression trees (CART)algorithm. For regression in CART, there is response vector Y which represents the response values for each observation in variable matrix X. The matrix X and vector Y can be randomly split into different subsets to regress trees. In each terminal nodes of the tree, a simple and accurate model can be constructed to explain the relationship of X and Y in this node. Amongst many non-parametric regression approaches, the RF is receiving considerable attention of ecological and other applications. Finally, monthly, seasonal and annual calibrated TRMM 3B43 precipitation data over the Tibetan Plateau during 1998-2012 were obtained. The results demonstrate that the accuracy of calibrated TRMM 3B43 precipitation data increased significantly. The maximum values of R Square(R²) of monthly average TRMM 3B43 precipitation data are in March and October, and reach to 0.9. Minimum R Square is 0.5 in December and January, which may be due to that the low precipitation may be more sensitive to the error. All of coefficient of efficiency(E)of monthly average calibrated TRMM 3B43 precipitation data is positive. The maximum value of it is 90. It shows that the calibrated precipitation can be used in research already. The minimum R square of seasonal and annual average calibrated TRMM 3B43 precipitation data is 0.58. Monthly and seasonal results relating to the first quarter, January, February and December, are not as good as others. The method should be improved to calibrate these data more efficiently in future.