An Uncertainty-Based Outlier Detection Method for Satellite-Derived Land Surface Temperature Validation Using In Situ Measurements

Land surface temperature (LST) is a crucial parameter driving water and heat exchange at the surface−atmosphere interface. Satellite-derived LST requires rigorous validation to ensure its reliability in Earth system modeling and climate change research. To address validation accuracy degradation caused by cloud contamination artifacts and satellite−ground spatiotemporal mismatch errors, conventional mean- and median-based outlier detection methods were commonly used in the validation of satellite-derived LST products using in situ measurements. However, both methods are based solely on the degree of deviation within statistical data itself, without considering the uncertainties associated with satellite-derived and ground-based LST. This limitation could result in the biased identification of outliers in satellite-derived LST validation. In this study, an uncertainty-based method was proposed to detect outliers in the validation of Moderate Resolution Imaging Spectroradiometer (MODIS)-derived LST using in situ measurements. This method quantifies total LST uncertainty budgets to flag anomalous data points by integrating uncertainties from both satellite retrievals and ground observations. Validation results across surface radiation budget network (SURFRAD) sites demonstrate the method’s efficacy when compared with those without outlier detection. Daytime implementation achieves significant root-mean-squared error (RMSE) reductions, notably at the Bondville (BND) site with a 3.1-K improvement, while nighttime applications yield marginal enhancements (<0.4 K), reflecting diminished thermal contrast and uncertainty components during nighttime. The uncertainty-based method consistently outperforms conventional mean- and median-based methods during daytime, with RMSE improvements ranging from 0.2 K at desert rock (DRA) to 2.6 K at BND. Site-specific variations highlight the method’s sensitivity to surface heterogeneity and vegetation dynamics. All methods exhibit comparable performance at night ( $\Delta $ RMSE < 0.15 K).