PM2.5浓度空间估算的神经网络与克里格方法对比.

Performance of artificial neural network modeling and Kriging interpolation in PM2.5 concentration estimation varies with sample sizes and predictor variables change. This paper analyzes the performance of ordinary Kriging (OK), radical basis function (RBF) networks based on geographic coordinates, CoKriging and RBF with the key factor(s) (CK and CoRBF) selected by correlation analysis and RBF network, using different training sets with various sizes. The spatial distribution of PM2.5 concentration is then estimated by the best performed method. Results show that RBF, CoRBF, OK, and CK can all be used to estimate PM2.5 concentration efficiently, and their accuracies improved unstably as the number of training sites increase. CoRBF with the key factor of population illustrates the largest variation of PM2.5 concentration, while CK has the highest coefficient of determination (R² ) and index of agreement (IOA) and the lowest mean square error (MSE), mean absolute error (MAE), and relative error (RE). Correspondingly, the spatial pattern of CK estimated PM2.5 concentration is smoother than CoRBF estimated PM2.5 concentration, while they both are very similar to site measurements and reveal detailed information. [ABSTRACT FROM AUTHOR]