机器学习模拟热环境及热岛时空变化特征研究

Simulation of urban thermal environment based on machine learning and research on the temporal and spatial variation characteristics of urban heat island

针对城市热岛研究中MODIS地表温度数据云覆盖像元值问题,该文构建SVR、RF和XGBoost 3种机器学习模型以实现对MODIS地表温度产品缺失像元值的最优模拟,以最优模型获得郑州市2009—2018年无缝地表温度数据;基于距离加权不透水面聚集密度,提取出郑州市逐年主要建成区范围;结合地表温度数据、建成区范围及郊区范围动态分析郑州市热岛面积及热岛强度变化。结果表明:XGBoost模型回归拟合系数为0.95,均方根误差为0.06,填补MODIS地表温度数据像元值空缺效果优于SVR和RF模型;研究期间内,郑州市每年主要建成区面积增长速率均超过8%;郑州市这10 a间地表温度的标准差呈平稳下降的趋势,中温区面积呈上升趋势,热岛强度呈下降趋势,表明了郑州市“高温化”的缓解。

Aiming at the problem of cloud coverage pixel value of MODIS Land Surface Temperature in urban heat island research.In this paper,three machine learning models,namely,Support Vector Regression(SVR),Random Forest(RF),and XGBoost,are implemented to achieve optimal simulation of missing pixel values for MODIS Land Surface Temperature,and the optimal model is used to obtain seamless surface temperature data in Zhengzhou from 2009 to 2018;Based on the distance-weighted cohesion density of the impervious surface,the annual built-up areas in Zhengzhou are extracted;Combined with surface temperature data,built-up area and suburban area,is used to dynamically analyze the urban heat island area and intensity changes of Zhengzhou.The results show that the XGBoost model’s coefficient of determination is 0.95,its root mean square error is 0.06,and its effect of filling the vacancy of the MODIS Land Surface Temperature data pixel value is better than that of the SVR and RF models.The annual growth rate of the main built-up areas exceeds 8%in Zhengzhou during this study period.The standard deviation of surface temperature in Zhengzhou shows a steady downward trend,the area of medium-temperature areas an upward trend,and the urban heat island intensity a downward trend,indicating the relief of Zhengzhou’s“high temperature”.