A New Flexible Approach for Reconstructing Satellite-Based Land Surface Temperature Images

Abstract

Time series of spatially continuous satellite data are increasingly used for environmental studies. Among these, land surface temperature (LST), retrieved from data such as the MODerate resolution Imaging Spectroradiometer (MODIS), plays a vital role in numerous applications. However, cloud cover significantly reduces the number of usable pixelwise LST observations. Despite various documented methods for reconstructing missing LST pixels, challenges remain regarding their flexibility to handle varying gap percentages and reliance on multiple ancillary datasets. This study presents a flexible and automated technique to reconstruct missing LST pixels without relying on ancillary data. The approach combines three innovative techniques: global regression analysis, local regression analysis, and geospatial analysis. The missing pixels percentage of each day determines the appropriate technique to fill the gaps. The method was applied to daily Terra MODIS LST datasets (MOD11A1) at 1 km spatial resolution from 2002 to 2022. Two evaluation methods were conducted: comparing with in-situ measurements and introducing artificial gaps. The validation was demonstrated over the Heihe River basin in China and in four experimental areas worldwide with available ground measurements from FLUXNET. Validation with artificial gaps produced average root-mean-square error (RMSE) and mean absolute error (MAE) of 2.33 K and 1.76 K, respectively. In-situ measurements indicated superior performance with R
², RMSE, and MAE of 0.85, 4 K, and 3.4 K, outperforming two existing methods. The study demonstrates that the model accurately reconstructs missing pixels on heterogeneous surfaces under diverse conditions, effectively handling large datasets and complex gaps.