An Effective Pixel Decomposition Approach For Land Surface Temperature Of Landsat 8

Land Surface Temperature (LST) is not only an important indicator to characterize the surface energy balance, but also one key parameter in the physics of land surface processes from local through global scales, which reflects surface properties. The spatial and temporal distribution of LST has certainly changed significantly because of the complexity and diversity of the surface composition. LST image with precise estimates is very important for application to many studies such as urban heat island monitoring, agricultural drought monitoring, soil moisture evaluation and global climate change. LST image with high spatial resolution compatible with thermal radiation of land surface components has always been an important direction of thermal infrared remote sensing, which is only the best means to fast access the regional or global land surface temperature.LST images retrieved from the thermal infrared (TIR) band data have much lower spatial resolution than its visible and near-infrared (VNIR) band data in main current remote sensing platform equipped with infrared sensors, which have limited their capability in applying to many studies required high spatial resolution. How to integrate the TIR band data with high spatial resolution of VNIR band data has been an important way to obtain high spatial resolution LST data.As to Landsat 8, the LST retrieved from TIR band data is with a spatial resolution of 100m under nadir viewing while VNIR band data are with a spatial resolution of 30m. The objective of this paper is to develop an effective pixel decomposition approach for land surface temperature to increase the spatial resolution with 30m using the VNIR band data as assistance based on Landsat 8.In the view of the conservation of energy, an attempt has been made in the paper to develop an algorithm for LST decomposition in the basis of the thermal radiation under different surface composition, which can be termed Temperature Unmixing with Spectral (TUS). Firstly, land surface fraction was obtained based on the Linear Spectral Mixing Model (LSMM) and land surface temperature of typical endmember was selected through Temperature Vegetation Index (TVX), so pixel decomposition of LST can be achieved integrated emissivity with different surface components. As for the law of energy conservation, we call the initial temperature estimation for pixel decomposition with a spatial resolution of 30 m. Secondly, the thermal radiance and thermal flux can be estimated for each sub-pixel of the initial temperature estimation, which can be used to the determination of radiance weights for the sub-pixels of the decomposed resulted image. Finally, the pixel decomposition of LST with high spatial resolution of 30m can be calculated based on the radiance weights for the sub-pixels.In order to prove the feasibility of our approach, we used two methods to verify the analysis and compare the accuracy. The one is the upscale-downscaled verification method. We scale the LST with the resolution of 100m up to 300m, then downscaled to 100m spatial resolution with TUS algorithm. Contrast to the original LST retrieved from improved mono-window algorithm (IMWA), TUS can effectively improve the spatial resolution of land surface temperature, reflecting the spatial differences of surface components. The other one is the comparison method to SUTM and E-DisTrad models. TUS can effectively ensure the energy balance before and after pixel decomposition, with the highest accuracy of MAE and RMSE are 1.71K and 2.37K, for MAE and RMSE of SUTM and E-DisTrad models are 2.23K,3.20K and 2.04K,2.79K, respectively. Therefore we can conclude that our TUS model is applicable for decomposition of LST images for high spatial resolution.The innovation point of this paper is to develop an effective approach for pixel decomposition of land surface temperature based on Landsat 8 with the conservation of energy. The feasibility in the complex surface coverage area has also been discussed, which has a better decomposition result with the lowest error.