Research On Spatial Downscaling Method Of Land Surface Temperature And Its Application

Land surface temperature(LST)is an essential physical parameter of land surface energy flux.Land surface temperature data is the basic data of various environmental studies.However,the low spatial resolution of existing land surface temperature products will lead to thermal mixing effect,which limits the application of land surface temperature data.Previous studies have shown that spatial downscaling of land surface temperature is a feasible and effective technology to solve this problem.Therefore,the study on spatial downscaling of land surface temperature has great theoretical meaning and application value.Previous studies did not consider both nonlinearity and spatial nonstationarity between land surface temperature and its influencing factors(such as land cover type,soil moisture and topography etc.),resulting in large uncertainty of downscaled results.In order to solve the above problems,this paper proposes a spatial downscaling method of land surface temperature based on multi-factor geographically weighted machine learning(MFGWML)algorithm.This paper takes Beijing City as the study area,and explores MFGWML model and its universality and application.The main research contents and conclusions are as follows:(1)This paper proposes an objective method to select the optimal feature combination.By analyzing the correlation between features and land surface temperature and the correlation between features,the multi-dimensional feature variables are preliminarily screened,and the optimal feature combination of each base model is determined by combining the results of variable importance provided by each base model.This objective feature selection method is confirmed effective,it can reduce the number of input variables greatly,thus reducing the complexity of the model,decreasing the running time and memory consumption,and avoiding overfitting.(2)This paper constructs a multi-factor geographically weighted machine learning model for spatial downscaling of land surface temperature.Based on Landsat 8 and Sentinel-2A images,this paper studies the spatial downscaling of land surface temperature to obtain land surface temperature data with high spatial resolution(10 m).The results show that: compared with the classical single-factor algorithm(the thermal image sharpening,Ts HARP),the RMSE values of MFGWML model are decreased by55.452% ? 58.949% under six downscaling schemes;compared with the classical twofactor model(the high resolution urban thermal sharper,HUTS),the RMSE values of MFGWML model are decreased by 43.782% ? 50.389% under six downscaling schemes.MFGWML combines the advantages of multiple regression model,machine learning model and geographically weighted model so that it can better identify the local spatial heterogeneity of land surface temperature and generate more accurate,reliable and robust spatial downscaled land surface temperature.(3)This paper explores the applicability of MFGWML model to domestic Gaofen satellite images.Using GF-1 WFV2 image,GF-6 PMS image and GF-2 PMS image as test data,the applicability of MFGWML model to domestic Gaofen satellite images are evaluated.The results show that: in the results of spatial downscaling of land surface temperature based on GF-1 WFV2 image,compared with the results of multi-factor geographically weighted regression(MFGWR)model and Ts HARP model,the RMSE values of MFGWML model are reduced by 12.889% and 74.211%,respectively.In the results of spatial downscaling of land surface temperature based on GF-6 PMS image,compared with the results of MFGWR model and Ts HARP model,the RMSE values of MFGWML model are decreased by 46.571% and 66.123%,respectively.In the results of spatial downscaling of land surface temperature based on GF-2 PMS image,compared with MFGWR model and Ts HARP model,the RMSE values of MFGWML model are decreased by 3.768% and 26.386%,respectively.In addition,the downscaled results of MFGWML model based on GF-1 WFV2 image,GF-6 PMS image and GF-2PMS image have high accuracy,and the RMSE values are 0.980 K,0.561 K and 0.664 K,respectively,which indicates that MFGWML model has good applicability to Gaofen satellite images.(4)This paper studies the application of the spatial downscaled land surface temperature data.Firstly,an object-oriented classification method integrating random forest algorithm and support vector machine algorithm is proposed to achieve highprecision land use classification of Sentinel-2A image.Secondly,with the support of 10 m LST data downscaled based on Sentinel-2A image and the remote sensing based ecological index,the impact of different land use types on the urban eco-environmental quality in Beijing is quantitatively analyzed.The results show that the urban ecoenvironmental quality is positively correlated with the proportion of vegetation,and negatively correlated with the proportion of impervious surface.Therefore,increasing vegetation coverage or reducing impervious surface can help to improve the urban ecoenvironmental quality.In addition,the impact of land surface temperature data before and after spatial downscaling on the evaluation of urban eco-environmental quality and urban heat island effect is compared and analyzed.The results show that the RSEI image and surface heat field information based on spatial downscaling 10 m LST are more refined than that based on resampling 10 m LST,indicating that the spatial downscaling of land surface temperature has obvious advantages.