基于地理加权回归模型评估土地利用对地表水质的影响

针对传统线性回归模型大多忽视空间数据局部变化特征这一缺陷,引入地理加权回归模型(GWR)用于评估土地利用对地表水质的影响,分析了不同子流域内两者关系出现空间变化的规律并阐释了原因.同时,对比了GWR模型与普通最小二乘模型(OLS)的校正R2、Akaike信息准则(AICc)及残差的空间自相关指数(Moran's I),验证了GWR模型在预测精度和处理空间自相关过程中是否优于OLS模型.结果表明,同一土地利用类型对水质的影响随空间位置的改变而发生方向或大小的变化.以温瑞塘河流域总氮(TN)与农用地的关系为例,从GWR模型局部回归系数的方向分析,两者关系表现为农村正、城区负的现象,从大小分析,旧城区TN与农用地回归系数的绝对值高于其它区域;在溶解氧(DO)与人口密度所构建的GWR模型中,两者关系在整个研究区域内均表现为负值,与OLS结果吻合,从回归系数的大小分析,人口密度对DO的作用在郊区及农村更为显著.针对此类关系出现空间变化的原因分析表明,相邻子流域土地利用百分比的改变及水体主要污染源的不同,是导致土地利用对水质作用发生变化的根本因素.最后,对比所构建的80个GWR与OLS模型校正R2、AICc指标,验证了GWR作为一种局部统计模型,其预测精度优于OLS等传统全局模型且更能反映实际空间特征.

Evaluating the impacts of land use on surface water quality using geographically weighted regression

Most traditional linear regression models ignore local variations of spatial data. In this study, a new technique called geographically weighted regression model (GWR) was introduced to evaluate the impacts of land use on surface water. The reason for the spatial variations of relationships between land use and water quality were explored. Meanwhile, the adjusted R², Akaike information criterion (AIC_c) and spatial autocorrelation index (Moran's I) of residuals were compared with ordinary least squares model (OLS) to verify if the GWR model is better than OLS in the prediction accuracy and the capacity of conducting spatial autocorrelation. The results showed that impact of the same types of land use on water quality changes in direction or size along with the variation of spatial position. For example, the relationships between TN and agricultural land in Wen-Rui Tang River showed a positive correlation in countryside and negative correlation in urban area in GWR models. The absolute values of regression coefficients in old downtown area were higher than other places. In the GWR model of dissolved oxygen (DO) and population density, the relationships were negative in the whole study area, which was consistent with the OLS results, and the effect of population density on DO is greater in suburban and rural areas. The reason for these spatial changes over different sub-watersheds indicated that the changes of land use percentage and the varied main pollution sources are fundamental factors. Furthermore, the adjusted R² and AIC_c values from the 80 established models confirmed that as a local statistical model, GWR had better prediction accuracy than OLS model and could better reflected the actual spatial characteristics.