Investigation Of Optimal Moving Window Size And Evaluation Index For Disaggregation Of Land Surface Temperatures

Disaggregation of land surface temperature(DLST),which aims at the generation of LSTs with fine resolutions,has been drawing increasing attention since 1980s.The past three decades have been witnessing the emergence of DLST methods in large numbers,making this field steadily evolve as a relatively independent subfield of thermal remote sensing.So far,this field has established relatively complete theoretical system and methodology.However,in terms of methodology,researchers have not explored much on moving window size,which directly influence the performance of regression for DLST.Also,evaluation system has been neglected even to now.These two aspects may seriously restrict the development of DLST in the future.Focusing on these two issues,this paper made a deep exploration and provided two coping strategies.To balance disaggregation accuracy and computational complexity,we found that the optimal moving-window size(MWS)for the LWS could be estimated by its statistical linear relationship with the resolution ratio between pre-and post-disaggregated LSTs.The results show that the selection of the better strategy is difficult and it generally depends on the associated land cover status.We therefore formulated an indirect criterion based on aggregation-disaggregation(ICAD),which instructs the selection of the better strategy between the LWS and GWS from median-to high-resolution by their performances from low-to median-resolution.Validations illustrate that the accuracy predicted by the ICAD is 72%.In cases when predictions are incorrect,there is no great difference between the GWS and LWS.Further analyses indicate that the LWS can be further improved by using a locally varying MWS,which requires the availability of historical high-resolution LSTs.To better assess DLST method performances under diversified scenarios,we formulated five protocols,through which a simple yet flexible index(SIFI)was subsequently designed.The establishment of SIFI includes four steps:(1)a detail-based evaluation,designed mainly to exclude the impacts of systematic deviations on estimated LSTs,(2)a Gaussian normalization,primarily intended to remove the differences in temperature unit and thermal contrast,(3)a triple comparison,aiming at attenuating the influence of the difference in resolution ratio on comparisons of method performances,and(4)a piecewise comparison,mainly scheduled to distinguish the three sharpening status including the under-sharpening,acceptable over-sharpening,and unacceptable over-sharpening.The ability of SIFI on evaluation was compared with those of root mean squared error(RMSE),Erreur Relative Globale Adimensionnelle de Synthese(ERGAS),and image quality index(Q)using simulation tests and real thermal data.The results illustrate that SIFI generally outperforms the other indexes:It is able to mitigate the impacts from process errors and controls on evaluation as well as able to indicate the sharpening status accurately.In conclusion,the strategies to determine the optimal moving window size and a new evaluation index have been achieved to help further improve the performance of DLST methods.