Estimation of 30 m land surface temperatures over the entire Tibetan Plateau based on Landsat-7 ETM+ data and machine learning methods

Abstract

Land surface temperature (LST) is an important parameter in land surface processes. Improving the accuracy of LST retrieval over the entire Tibetan Plateau (TP) using satellite images with high spatial resolution is an important and essential issue for studies of climate change on the TP. In this study, a random forest regression (RFR) model based on different land cover types and an improved generalized single-channel (SC) algorithm based on linear regression (LR) were proposed. Plateau-scale LST products with a 30 m spatial resolution from 2006 to 2017 were derived by 109,978 Landsat 7 Enhanced Thematic Mapper Plus images and the application of the Google Earth Engine. Validation between LST results obtained from different algorithms and in situ measurements from Tibetan observation and research platform showed that the root mean square errors of the LST results retrieved by the RFR and LR models were 1.890 and 2.767 K, respectively, which were smaller than that of the MODIS product (3.625 K) and the original SC method (5.836 K).

Keywords

• google earth engine
• remote sensing
• machine learning
• land surface temperature
• random forest