基于随机森林的陕西省西安市近地表气温估算

Estimation of Near-surface Air Temperature in Xi'an City of Shaanxi Province,China Based on Random Forest

随着城市规模不断扩大以及人口激增,城市气候与热环境问题日益凸显,开展城市近地表气温遥感监测研究能够为改善城市气候、减缓热岛效应、打造适宜人居环境提供参考。针对传统气温监测方法在多因素复杂关系模拟中的局限性,以陕西省西安市为研究对象,运用可以集成多要素、学习复杂、非线性映射关系的随机森林(Random Forest,RF)模型,基于Landsat 8卫星遥感数据以及SRTM高程数据相关参数的综合分析,构建多种近地表气温估算的随机森林模型,通过性能对比评估优选最佳模型,估算了2016年5月16日西安市近地表气温,分析了近地表气温的空间分布特征。结果表明:在所有近地表气温影响因子中,高程对随机森林模型近地表气温估算的贡献度最大,其次是地表温度。所有随机森林模型训练集的判定系数(R^(2))均高于0.916,均方根误差(RMSE)均低于0.467℃,验证集判定系数均高于0.726,均方根误差均低于0.840℃;训练集判定系数均高于验证集,均方根误差均低于验证集;最优随机森林模型训练集判定系数为0.934,均方根误差为0.425℃,验证集判定系数为0.795,均方根误差为0.783℃;气温估算精度判定系数为0.792,均方根误差为1.055℃。西安市中心城区气温高于郊县区,中心城区最低气温平均值、最高气温平均值及气温平均值分别高于郊县区1.54℃、0.01℃和1.76℃。综上所述,西安市近地表气温南低北高,空间差异明显,自中心城区、郊县区至南部山区逐渐降低,呈现出显著的城市热岛效应。

With the continuously expanding of the urban areas and the dramatic population growth,serious urban climate and thermal environment issues have occurred and attracted extensive attention.High-precision and large-scale remote sensing monitoring of near-surface air temperature in Xi'an city can provide reference for the improvement of urban climate,the alleviation of heat island effect,as well as the creation of a suitable living environment.Traditional near-surface air temperature monitoring approaches are confined when dealing with complicated relationships among multiple factors.Taking Xi'an city as the study area,the random forest(RF)models,which can integrate multiple factors and simulate complicated and non-linear mapping relationship,were adopted.Through comprehensive analysis of various parameters derived from Landsat 8 imagery and SRTM DEM data,several near-surface air temperature estimation RF models were constructed,and the optimal model was confirmed by performance comparison and evaluation.Furthermore,near-surface air temperature in Xi'an city on May 16,2016 was estimated,and the spatial distribution pattern was discussed.The results show that among all the near-surface air temperature influencing factors,altitude has the highest contribution to the estimated temperature of the RF models,followed by land surface temperature(LST);both of them are crucial for the estimation of near-surface air temperature.For all the RF models,the training set R^(2) is higher than 0.916,and the RMSE is lower than 0.467℃;all the validation set R^(2) is higher than 0.726,and the RMSE is lower than 0.840℃;the training set R^(2) of all RF models is higher than that of the validation set,while the training set RMSE is lower compared with the validation set;the training set R^(2) of the optimal model is 0.934 with the RMSE being 0.425℃,and the validation set R^(2) is 0.795 with a RMSE value of 0.783℃;the R^(2) and RMSE of the air temperature estimation accuracy are 0.792 and 1.055℃,respectively.Near-surface air temperature in the central urban districts of Xi'an city is higher than that of suburban areas,and the mean minimum temperature,the mean maximum temperature,as well as the mean temperature are higher than those of suburban areas by 1.54℃,0.01℃and 1.76℃,respectively.The near-surface air temperature in the study area is low in the south and high in the north,and presents obvious spatial heterogeneity.Air temperature gradually decreases from the central urban districts through suburbs to the southern mountainous areas,indicating a prominent urban heat island effect.