Downscaling Of Urban Surface Temperature In Typical Karst Mountainous Areas

The land surface temperature(LST)is a very important environmental factor and intermediate parameter in ecological environment monitoring.However,the spatial resolution of thermal infrared data is limited by the factors,such as its long wavelength and low energy,LST often cannot meet the application requirements.In karst mountain city with fragile ecological environment,due to the large terrain undulations and the high fragmentation of the landscape,it is urgent need for LST with high spatial resolution to detect the difference of the broken underlay surface.This paper chooses Guiyang City,a typical karst mountain city,as the research object,and two areas dominated by natural surface and construction land are selected as the study areas.Based on Landsat-8 TIRS data,Sentinel-2 data and ALOS PALSAR data,etc,through the LST inversion and machine learning method to estimate the high spatial LST of the karst mountain city,establish a LST downscaling method suitable for karst mountain city,and obtain LST products with 10m spatial resolution,then the correlations between terrain factors and LST were analyzed in order to reduce temperature differences caused by terrain effects.The main research results are shown as follows:(1)In the land surface temperature inversion of Landsat-8 thermal infrared image,the retrieval results based on single-channel algorithm and single-window algorithm,although the overall trends are generally consistent in two study areas,the single-window algorithm achieves good results at the actual measurement site,whose average absolute error is 2.81K.(2)In the process of building a more suitable downscaling model in karst mountain city,taking the karst topography and geomorphology characteristics of the research areas into consideration,the scale factors are screened,and different research areas show different results caused by different underlying surface characteristics.The scale factors in the study area dominated by natural surface are:short-wave infrared band(MSI11),vegetation red edge band(MSI5),short-wave infrared band(MSI12),green band(MSI3),red band(MSI4),blue band(MSI2),sky view factor(SVF),fraction vegetation(FV),vegetation ratio index(RVI),normalized vegetation index(NDVI),bare soil index(BSI),vegetation red edge band(MSI6),digital elevation model(DEM),index-based built-up index(IBI),hillshade,normalized difference build-up index(NDBI),vegetation red edge band(MSI7),vegetation red edge band(MSI8A),near-infrared band(MSI8);The scale factors in the study area dominated by construction land are:short-wave infrared band(MSI12),blue band(MSI2),red band(MSI4),green band(MSI3),index-based built-up index(IBI),vegetation red edge band(MSI5),ratio vegetation index(RVI),normalized multi-band drought index(NMDI),fraction vegetation(FV),normalized difference vegetation index(NDVI),bare soil index(BSI),short-wave infrared band(MSI11),normalized difference build-up index(NDBI),soil adjusted vegetation index(SAVI),relief degree of land surface(RDLS),slope,solar incident angle(SIA).(3)In the downscaling of LST in the karst mountain city,a comparison with three downscaling methods,such as the image thermal sharpening algorithm(Ts HARP),the random forest model(RF)and the e Xtreme gradient boosting model(XGBoost),was made.For these three algorithms,the XGBoost model are the best.On the accuracy evaluation of ground temperature measurement,the average absolute error of the XGBoost model is 1.67K,followed by RF model with an average error of 1.90K,and Tsharp algorithm is the worst with an average error of 2.41K.On the accuracy evaluation of the upscaling,the MAE,RMSE and R~2 of XGBoost model in the study area dominated by natural surface are 0.51k,0.67k and 0.92respectively,while those in the study area dominated by construction land are 0.58k,0.81k and0.91 respectively.It shows that the XGBoost model can capture the more important scale factors that affect the LST by assigning weights to each weak estimator,so as to obtain higher-precision results.(4)In the terrain suppression of LST,it is suitable to use the solution of aspect-based partition and DEM correction in both study areas.The relationship between DEM and LST is constructed in the two study areas dominated by natural surface and construction land,and the coefficient coefficient of the linear equation are 0.831 and 0.836 respectively.Through this relationship,the surface temperature of the study area dominated by natural surface and the study area dominated by construction land will be reduced by 2K and 1K respectively,making the topographic effect of LST in two study area is effectively suppressed.