耦合地理加权与随机森林的地表温度降尺度

Land Surface Temperature Downscaling in Coupled Geographically Weighted and Random Forests

针对当前主流地理加权地表温度降尺度算法仅考虑地表温度(land surface temperature,LST)与尺度因子间线性或简单非线性关系的问题,提出了一种利用随机森林表征复杂非线性关系并与GWR耦合(geographically weighted random forest model,GWRF)的LST降尺度框架。GWRF降尺度框架从反射率、光谱指数、地形因子等多种尺度因子中筛选出最佳因子,利用地理加权随机森林方法建立LST与尺度因子间复杂的局部非线性关系,实现1000 m LST降尺度到100 m。以北京和张掖地区作为实验区,并与地理加权回归、非线性地理加权回归、随机森林回归模型进行比较。研究发现,基于GWRF的降尺度模型在所有研究区均表现良好,均方根误差和平均绝对误差均低于其他模型,并且具有更高的决定系数R 2。

A land surface temperature(LST)downscaling framework using random forest to characterise complex nonlinear relationships and coupled with GWR(geographically weighted random forest model,GWRF)is proposed to address the problem that the mainstream geographically weighted LST downscaling algorithm only considers linear or simple nonlinear relationships with scale factors.The GWRF downscaling framework selects the best factors from reflectance,spectral index,topographic factors,etc.,and establishes the complex local nonlinear relationship between LST and scale factors using the geographically weighted random forest method,so as to achieve the downscaling of LST from 1000 m to 100 m.In this study,the experimental areas in Beijing and Zhangye are used as the experimental areas,and the geographically weighted random forest model is used as the experimental area to determine the downscaling of the LST.Compared with the geographically weighted regression(GWR),nonlinear geographically weighted regression(NGWR),and random forest regression(RF)models,it is found that the GWRF-based downscaling model performs well in all the study areas,with the root-mean-square error(RMSE)and mean-absolute error(MAE)lower than those of the other models,and it also has a higher coefficient of determination,R 2.