中文摘要：

气温是大量农业、水文、气候、生态模型的输入变量.在地形复杂的区域,考虑气温与环境变量的线性回归关系和残差的自相关性的方法（如回归克里格法,regression Kriging,RK）是目前气温插值的主要方法.但此类方法多使用基于普通最小二乘的全局回归技术,没有顾及回归关系的空间非平稳性.地理加权回归克里格（geographically weighted regression-Kriging,GWRK）是一种既能顾及回归关系的空间非平稳性、又能考虑残差的自相关性的一种插值方法.本文用RK和GWRK对海南岛2013年12月18日的日平均气温进行插值并进行比较研究.依相关性分析和逐步回归分析的结果,采用RK1（以海拔为辅助变量）、GWRK1（以海拔为辅助变量）、RK2（以纬度、海拔、海陆距离为辅助变量）和GWRK2（以海拔、海陆距离为辅助变量）4种模型进行研究,并用80个验证站评估4种模型的精度.结果表明：GWRK1模型的最大正误差、最大负误差、平均绝对误差、均方根误差均最接近于0.从最大正误差、平均绝对误差、均方根误差3个指标看,考虑更多辅助变量的RK2、GWRK2模型反而不及只考虑海拔的RK1、GWRK1模型,表明RK2、GWRK2模型中辅助变量之间的相关性对插值结果有较大影响.

英文摘要：

Air temperature is the input variable of numerous models in agriculture,hydrology,climate,and ecology. Currently,in study areas where the terrain is complex,methods taking into account correlation between temperature and environment variables and autocorrelation of regression residual（ e. g.,regression Kriging,RK） are mainly adopted to interpolate the temperature. However,such methods are based on the global ordinary least squares（ OLS） regression technique,without taking into account the spatial nonstationary relationship of environment variables. Geographically weighted regression-Kriging（ GWRK） is a kind of method that takes into account spatial nonstationarity relationship of environment variables and spatial autocorrelation of regression residuals of environment variables. In this study,according to the results of correlation and stepwise regression analysis,RK1（ covariates only included altitude）,GWRK1（ covariates only included altitude）,RK2（ covariates included latitude,altitude and closest distance to the seaside） and GWRK2（ covariates included altitude and closest distance to the seaside） were compared to predict the spatial distribution of mean daily air temperature on Hainan Island on December 18,2013. The prediction accuracy was assessed using the maximum positive error,maximum negative error,mean absolute error and root mean squared error based on the 80 validation sites. The results showed that GWRK1’s four assessment indices were all closest to 0. The fact that RK2 and GWRK2 were worse than RK1 and GWRK1 implied that correlation among covariates reduced model performance.