Correction of Overestimation in Observed Land Surface Temperatures Based on Machine Learning Models

Abstract

Abstract Land surface temperature (LST) is an essential variable for high-temperature prediction, drought monitoring, climate, and ecological environment research. Several recent studies reported that LST observations in China warmed much faster than surface air temperature (SAT), especially after 2002. Here we found that the abrupt change in daily LST was mainly due to the overestimation of LST values from the automatic recording thermometer under snow cover conditions. These inhomogeneity issues in LST data could result in wrong conclusions without appropriate correction. To address these issues, we proposed three machine learning models—multivariate adaptive regression spline (MARS), random forest (RF), and a novel simple tree-based method named extreme gradient boosting (XGBoost)—for accurate prediction of daily LST using conventional meteorological data. Daily air temperature (maximum, minimum, mean), sunshine duration, precipitation, wind speed, relative humidity, daily solar radiation, and diurnal temperature range of 2185 stations over 1971–2002 from four regions of China were used to train and test the models. The results showed that the machine learning models, particularly XGBoost, outperformed other models in estimating daily LST. Based on LST data corrected by the XGBoost model, the dramatic increase in LST disappeared. The long-term trend for the new LST was estimated to be 0.32° ± 0.03°C decade−1 over 1971–2019, which is close to the trend in SAT (0.30° ± 0.03°C decade−1). This study corrected the inhomogeneities of daily LST in China, indicating the strong potential of machine learning models for improving estimation of LST and other surface climatic factors.