Uncertainty estimation method and Landsat 7 global validation for the Landsat surface temperature product

Abstract

Surface temperature is a valuable metric for many Earth monitoring applications, which motivated the development of the Landsat Surface Temperature (LST) product. The initial LST algorithm, developed by Cook, was geographically restricted since the atmospheric inputs and truth data were limited to North America (Cook, 2014. Atmospheric Compensation for a Landsat Land Surface Temperature Product. Rochester Institute of Technology). The main objectives for this product are to produce an LST image for every available Landsat thermal image, and to also provide a per-pixel estimate of LST uncertainty. Various studies were performed in order to allow the LST algorithm to operate globally, after which a thorough global validation study was performed for Landsat 7 images. In this study, the LST algorithm was found to have an average error for all cases of −0.211 K compared to the MODIS Sea Surface Temperature (SST) product. For cases where transmission was less than 0.3 and clouds were within 1 km of the validation, the LST RMSE was 2.61 K. When transmission was at least 0.85 and clouds were more than 40 km away, the RMSE was 0.51 K. A LST uncertainty estimation method was developed that utilizes standard error propagation in combination with observed trends between LST error, transmission, and cloud proximity. When the uncertainty method was applied to the global validation dataset, 20% of the estimated LST uncertainties were less than 1 K and 63% were less than 2 K. LST uncertainty can be extended to other Landsat sensors pending small validation studies, and will be an extremely beneficial tool for users, since it allows them to select all pixels in an LST image that meet their accuracy requirements.