Estimation of land surface temperature from AMSR2 microwave brightness temperature using machine learning methods

Passive microwave has been used in land surface temperature (LST) inversion because of its all-weather availability. This paper proposed a LST retrieval method based on machine learning by integrating multiple datasets, including Advanced Microwave Scanning Radiometer 2 (AMSR2), Global Forecast System (GFS) reanalysis datasets, ERA5-land reanalysis products, Moderate Resolution Imaging Spectroradiometer (MODIS) products, Shuttle Radar Topography Mission (SRTM), and in situ measurements. Four algorithms, including linear regression (LR), random forests (RF), extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM), were explored and compared by ten-fold cross-validation. The best-performing LightGBM algorithm was applied to train the model, and day and night models were developed separately. The models were validated using in situ LST of training stations from 2017 to 2019. The day and night models root mean square error (RMSE) are 3.23 and 2.43 K, and that of bias are −0.24 and −0.36 K, respectively. The in situ LST from 2015 to 2019 that not used for model training were selected to further validate both day and night models, with RMSE = 5.42 and 2.91 K, and bias = −0.60 and 0.06 K, respectively. Validation results indicated that the temporal performance of the models is better than those of spatial performance and the night model performed better than the day model. Additionally, model performance in different seasons and land cover types demonstrated the robustness of the models over complicated surfaces. These results suggest that the LightGBM algorithm has good accuracy in LST estimation, making it possible to apply for the generation of LST at a global scale.