A novel method for optimizing air temperature estimation and quantifying canopy layer heat island intensity in eastern and central China

Abstract

Many predictors are considered in estimating the spatial distribution of air temperature. However, a key problem remains the optimal selection of predictors. In this study, genetic algorithm (GA) was used to select optimal regression variables among 17 predictors, and GA-selected predictors were applied to random forest (RF) regression models for mapping air temperature in eastern and central China. By comparison with observed air temperature, results indicate that estimation errors of RF regression models constructed with GA-selected predictors first decrease (up to 5 predictors) and then remain stable. It is found that optimal combination of regression variables has the lowest estimation error (RMSE = 1.21 °C) and the best fitting precision (R2 = 0.9662), including 5 predictors that are latitude, relative humidity, elevation, distance to prefecture city and distance to third-order stream. Next, we compared the quantification of canopy layer heat island intensity (CLHII) based on the analysis of observed and estimated air temperature. We infer that quantification of CLHII using ground monitoring network results in obvious deviation due to the sparse weather stations only providing limited information over wide areas. Finally, we evaluated the robustness in quantifying the CLHII based on estimated air temperature derived from GA-1, -2, -3, -4, -5 and -17 models. Results show that the estimation of air temperature with low estimation error facilitates quantifying accurate CLHII, emphasizing the importance of selecting optimal regression variables. In summary, results reveal the effectiveness of GA in selecting optimal regression variables and provide insight into quantifying CLHII.