A Comparative Study On Land Surface Temperature Retrieval Algorithms From Landsat-8 Imagery

Land Surface temperature(LST)is one of the most significant surface environmental parameters indicating the health of the earth and predicting future changes of the earth,and it can be considered as the “body temperature” of the earth.LST is closely related to various resource and environmental processes on the earth's surface,and plays an important role in human-environment interaction and global climate researches.Based on the importance of LST in the energy exchange and balance between the earth and the atmosphere,LST inversion has become an important topic in the field of quantitative remote sensing.Because of the inherent out-of-field stray light of Landsat-8,the absolute error of the radiation calibration of TIRS data incurs that the calibration accuracy still unable to meet the requirements.The split window algorithm used to reverse LST based on Landsat-8 data has always been a controversial topic owing to the prominent uncertainty of TIRS 11,which is doubled greater than that of TIRS 10.Until February 2017,the Landsat calibration team developed an out-of-field stray light algorithm to reduce the influence of TIRS bands on temperature inversion,which provided a broad prospect for the application of split-window algorithm in temperature inversion.As the most widely used remote sensing data for LST retrieval,a variety of LST retrieval algorithms based on Landsat-8 data,such as single-channel algorithm(SC algorithm)and several split-window algorithms(SW algorithms),have been proposed by previous researches.However,it has shown that the accuracy of SC algorithm and various SW algorithms has great inconsistencies among different regions.Therefore,aiming at these problems,the inversion accuracy of the generalized single-channel(GSC)proposed by Jiménez-Mu(?)oz et al.and five SW algorithms were compared and analyzed based on the in-situ measurements of seven NOAA's Surface Radiation Budget Network(SURFRAD)sites(PSU site,FPK site,BND site,TBL site,DRA site,SXF site,GWN site)with different land cover types and altitudes,and 72 high quality Landsat-8 imagery from January 1,2018 to June 30,2019.The work aims to provide an overall evaluation of the LST retrieval accuracy of GSC algorithm and the five SW algorithms.Furthermore,it will assist in the qualitative understanding of the effectiveness of the SLCA algorithm on the TIRS band.The main conclusions are as follows:1?From the perspective of sites,the retrieval accuracy of the six algorithms at the GWN site(pasture)and the PSU site(crop land)is the highest,and the algorithm performance is the best with the overall MAE values and RMSE values below 2.0 K and 2.5 K.The accuracy of the FPK site(grass land)comes second,and the overall MAE and RMSE values of all the algorithms are below 2.5 K and 3.0 K.The algorithm accuracy of the BND site(crop land)is slightly lower than that of the FPK site with overall MAE and RMSE below 3.0 K and 2.5 K.The accuracy of the SXF site(crop land)and the DRA site(shrub)is in the moderate level and the overall MAE values and RMSE values for both sites are below 3.5 K?3.5 K and 4.0 K?5.0 K,respectively.While,the accuracy over TBL site is the lowest and the overall MAE is below 10.0 K,and the overall RMSE is below 10.0 K.2?From the perspective of seasons,the inversion accuracy of the six algorithms shows a certain seasonal dependence.The comparison between the GSC algorithm and SW algorithms shows that: the inversion accuracy of the GSC algorithm is significantly higher than that of the five SW algorithms in spring.The possible reason is that the GSC algorithm uses a single thermal infrared band to perform LST inversion,which requires fewer input parameters and hence less uncertainty.While SW algorithms adopt both thermal infrared bands for LST inversion,then the required input parameters are more and the uncertainty factor increases correspondingly.In autumn,the accuracy of the GSC algorithm is lower than that of the SW algorithms.The possible reason lies in that when the atmospheric water vapor content is high,the SLCA algorithm is more effective.With the contribution to the recalibration of the two thermal infrared bands,the retrieval accuracy using that the combination of the two thermal infrared bands is much better than using only a single thermal infrared band.In summer,the accuracy of the GSC algorithm is higher than the SW-D algorithm,SW-J algorithm,and SW-R algorithm,but lower than that of the SW-Y algorithm and SW-JIM algorithm,which could be attributed to the dominate position of the input parameters of the SW-D algorithm,SW-J algorithm,and SW-R algorithm.The more input parameters are,the greater errors of the inversion result are.And when the atmospheric water vapor content is high,the effectiveness of the SLCA algorithm on the SW-Y and SW-JIM are particularly better represented.In winter,the accuracy of GSC algorithm is lower than that of SW-D algorithm,SW-Y algorithm,and SW-R algorithm,but higher than SW-J algorithm and SW-JIM algorithm.In general,the accuracy of the SW algorithm is higher.3?From the perspective of seasons,seasonal dependence was also obvious among the five split window algorithms.In spring and winter,the accuracy of the SW-R algorithm(about1.0 K)is the highest,and the accuracy of SW-Y and SW-JIM algorithm(<1.5 K)is slightly lower than that of SW-R algorithm.The accuracy of the SW-D algorithm(<2.0 K)is in the middle,and the accuracy of SW-J algorithm(between 5.0-6.0 K)is lower than the accuracy of other four SW algorithms.In summer and autumn,the accuracy of SW-JIM algorithm(<2.0 K)is the highest,the accuracy of SW-D algorithm,the SW-Y and SW-R algorithm(<3.0 K)are in the middle,and the accuracy of SW-J algorithm(5.5-6.5 K)is lower than the accuracy of other four SW algorithms.Generally speaking,the algorithm accuracy in spring and winter is better than in summer and autumn,and the accuracy of SW-JIM algorithm is better,the SW-R,SW-Y,and SW-D algorithms are moderate,and the accuracy of SW-J algorithm is slightly lower.4?From the correlation to in-situ measurements,LST retrieved from the six algorithms have a strong correlation with in-situ LST.The correlation coefficients of the six algorithms are 0.9837,0.9843,0.9701,0.9833,0.9835,and 0.9835,respectively,demonstrating that the inversion results have high consistency with in-situ LST.Among them,the correlation coefficient of SW-J algorithm is the lowest(0.9701),and the correlation coefficient of SW-D algorithm is the highest(0.9843).And GSC algorithm,SW-Y,SW-R and SW-JIM algorithm have close correlation coefficients(0.9837?0.9833?0.9835?0.9835).This shows that the introduction of the SLCA algorithm has preferably improved the data quality of the Landsat-8thermal infrared band,making the split-window algorithm have strong applicability in the LST inversion based on Landsat-8.5?From the perspective of key impact factors for LST retrieval,as far as atmospheric water vapor content is concerned firstly,the increase of atmospheric water vapor content will lead to decrease in atmospheric transmittance and increase the error of the LST inversion result of surface temperature.The summer is comparatively humid,and the atmospheric water vapor content is higher,and the atmospheric transmittance will decrease.Therefore,a greater inversion error occurs in summer,while lower in winter.Secondly,as regard to the emissivity,the vegetation in summer is denser,and the shadows projected on the ground are much heavier.However,it is the vegetation canopy,instead of the real surface,that is observed by the satellites,which will result in a larger surface emissivity error in summer compared with in winter.Thirdly,the input parameters and their combination complexity have an influence on the algorithm accuracy.The input parameters including brightness temperature,surface emissivity and atmospheric transmittance of the SW-J are the same as those of the SW-R algorithm.Due to more combinations of the three main input parameters and more intermediate variables to be calculated in the SW-J algorithm,its inversion accuracy is lower than the accuracy of other algorithms.Finally,the elevation of the site and the amount of data also have a certain impact on the inversion accuracy.The lowest elevation among the seven sites is at the GWN site(98m),and the highest is at the TBL site(1689m),and relatively poor accuracy and better accuracy are observed at the TBL site and GWN site,respectively.The elevation of the DRA site is as similar as TBL site,but the data volume of the DRA site(16scenes)is more than that of the TBL site(9 scenes).The inversion accuracy of the DRA site is higher than that of the TBL site,indicating that more convincing results can be obtained when more data involved.