降雨量地面观测数据空间探索与插值方法探讨

空间插值方法广泛应用于气象数据产品的制作,其精度与气象要素的空间变异特征、气象观测站分布和插值方法选择有关。选择美国得州599个地面观测站30年平均降雨量记录,设计了27个观测站样本方案,选择全年、1月和8月数据,利用空间统计、空间自相关、半变异函数等方法探索降雨量的空间变异特征,并采用5种常规方法进行空间插值,比较和解释插值结果,在此基础上讨论基于知识的气象要素空间插值方法。案例研究发现:①降雨量地面观测数据通常具有明显的空间趋势、较强的空间自相关特征和较稳定的空间变异规律,但针对不同时段或采样方案,其空间自相关强度和半变异函数模型会有一定的差异。②增加气象观测站数,空间插值误差有减小的趋势;但观测站数目达到一定数值后,增加观测站数,插值精度提高并不明显。③在观测站较少时,不同插值方法间的精度差异较大,而在观测站充足的情况下,其差异有减小的趋势。④探讨气象要素与地理环境要素之间的关系,获得定量化的先验知识,开发基于知识的空间插值模型,是高精度气象要素插值的关键;线性加权回归和地理加权回归方法的初步试验验证了这一思路的有效性。

Spatial exploration and interpolation of the surface precipitation data

Various spatial interpolation methods are widely applied to climate map production.The quality of climate spatial interpolation depends on the spatial variation of climate factors,the spatial distribution of climate stations,and the interpolation method.For examining the relationships between station distributions,interpolation methods and interpolation quality,599 climate stations in Texas,US with 30-year precipitation records are collected and 27 station samples are designed by regular or random sampling.The spatial patterns of Annual,January and August precipitation data are investigated using exploring spatial analysis such as spatial statistics,spatial autocorrelation testing,and semivariogram modeling.Five methods,i.e.,Kriging,IDW,local polynomial,regularized spline and thin plate spline,are used in the spatial interpolation of Annual,January and August precipitation data for all the station samples.The interpolation results,in terms of cross-validation errors,known-point check errors,and linear regression of the known values versus predicted values,are compared and discussed.Four findings are generalized from this case study.First,precipitation data usually have patterns such as obvious spatial trend,high-level spatial autocorrelation and stable semivariogram model.Nevertheless,the spatial patterns may vary by sample stations and seasonal changes.Considering these spatial characteristics,the exploring spatial data analysis is necessary and essential for climate spatial interpolation.Second,increasing the sample size of climate stations,the interpolation accuracy will be improved.But at a reasonable number of stations,increasing the sample size,the interpolation accuracy will not be improved obviously.Third,when the observation samples are scarce,different methods usually give very different interpolation results.When the samples are relatively rich,general methods tend to create similar results.Fourth,considering the intrinsic limitations of the general spatial interpolation methods,the authors suggest to explore the local relationships between climate factors and geographic variations,and to develop a knowledge-based interpolation method by introducing geographic variables and local regression models.The weighted linear regression of precipitation versus elevation for northwest Texas and the geographic weighted regression for entire Texas have shown the potentials of such new approaches.It is also argued that exploring spatial data analysis and knowledge-based spatial interpolation are critical for high-quality climate data interpolation.