稀疏样本下冬春季月平均气温空间插值研究——以新疆玛纳斯河流域为例

针对我国典型高寒山区——新疆天山中段玛纳斯河流域积雪-融雪过程模拟中气温空间数据的制备问题,以气象站点稀少的玛纳斯河流域为研究区域,利用最小二乘相关分析法开展了冬、春季(2015年11月-2016年4月)气温环境变量分析,通过共线性检测确定了纬度、海拔、坡度、坡向、NDVI 5个环境变量组成了最优因子集,构建了基于广义回归神经网络(GRNN)的月平均气温空间插值模型。采用区域内139个站点中的119个观测站点数据作为训练数据对GRNN模型进行训练,确定了冬、春季6个月的区域气温空间插值模型。利用剩余的20个观测站点数据作为检验样本,以均方根误差(RMSE)和平均相对误差(MRE)为评价指标,对模型的回归误差进行分析。结果表明:本模型6个月的平均RMSE值为1. 46,优于传统的地理加权回归克里金(GWRK)方法(其平均RMSE值为2. 22)。此外,从不同月份的气温空间插值分布图来看,本文模型空间插值后的气温变化趋势与实际变化趋势一致。从气温的空间分布情况来看,各空间点的气温与其海拔高程呈正相关,且随地表覆盖类型变化。这也表明本文提出的插值策略并组合建立的GRNN模型对于稀疏气象站点条件下的气温空间插值精度更高,一致性较好。

Spatial interpolation of monthly average air temperature during winter-spring season using sparse samples: A case study in Manas River Basin in Xinjiang

For preparation of air temperature spatial data used in the simulation of snow cover to snow melting process in the Manas River Basin with sparse meteorological stations in the middle part of Tianshan Mountains, Xinjiang, this paper carried out the analysis of environmental variables influencing air temperature in winter and spring (2015-11-2016-04) using least square correlation method so as to determine the optimal set of environmental variables through collinearity detection. Here the optimal factor set is composed of five environmental variables: latitude, elevation, NDVI, terrain slope and aspect. Then the Generalized Regression Neural Network (GRNN) model was constructed for spatial interpolation of monthly mean air temperature. Finally, the proposed GRNN model was trained with observation data of 119 stations among the total 139 in the region, and the spatial interpolation model of monthly air temperature for the six months in winter and spring were determined. Also the regression error of the proposed model was analyzed with Root Mean Square Error (RMSE) and the Mean Relative Error (MRE) as measures using the rest observation data of 20 stations as test samples. Results show that the averagely RMSE value of this model is 1.46 in six months, which is superior to the traditional GWRK method with an average RMSE value of 2.22. In addition, from the spatial interpolation maps of air temperature of 6 months, the varying trend of the interpolated air temperature using our model is consistent with that of actual circumstances. The temperature of each spatial site is positively correlated with its elevation, and varies with the type of surface coverage. In summary, the proposed GRNN model with combination of interpolation strategy has shown an improved interpolation accuracy and better spatial consistency for spatial interpolation of air temperature even if few meteorological observation stations were provided.