Estimation Of Land Surface Parameters And Actual Evapotranspiration Over The Tibetan Plateau Based On Machine Learning Methods

The Tibetan Plateau(TP),known as "the third pole",is the highest plateau in the world with an average altitude of more than 4000 m.The thermal and dynamic effects of the TP not only regulate the energy and water cycles in Asia,but also have a profound impact on the global climate and environment.Meanwhile,the plateau region has more than 50%of the total lake area and more than 80%of the total glacier area in China.Agricultural,industrial and domestic water for nearly 40%of the world’s population is provided by rivers originating from the TP.Therefore,the TP is also known as the "Asia water tower".An in-depth analysis of the ground-atmosphere interaction process on the TP is of great significance to the water resources and ecological security in the surrounding and downstream areas of the plateau.Land surface temperature(LST)and actual evapotranspiration(ETa)are key variables that can characterize land surface energy and water cycle processes,respectively.The accurate estimation of land surface parameters and ETa is essential to the recognition of climate change and the study of land-atmosphere water and heat exchange processes.Large-scale land surface temperature data can be obtained by satellite remote sensing.However,due to the complex terrain and unique climatic conditions on the plateau,current algorithms to retrieve land surface parameters often have large errors in the TP.Besides,due to the limitation of data temporal resolution,current research on the variation in ETa over the TP area can only be provided on monthly or annual scales.To obtain high spatial resolution land surface characteristic parameter datasets over the plateau area with high precision,an improved single-channel algorithm that is more suitable for the TP area was proposed in this study by using a linear regression model(LR)and in-situ measurements over the TP.The validation of the LR LST results showed that the retrieved LST has good agreement with in situ measurements with an RMSE of 2.767 K.Meanwhile,a new method to estimate the LST data using the random forest model(RF)based on different land cover types was also proposed in this study.The RF model showed the best accuracy in LST estimation with an RMSE of 1.890 K when compared with the MODIS LST product with an RMSE of 3.625 K and the USGS Landsat 7 Level-2 LST product with an RMSE of 4.493 K.Additionally,the LR and RF models trained by Landsat 7 ETM+data were applied to the two thermal infrared bands of the Landsat 8 Thermal Infrared Sensor(TIRS)to validate the portability of these models.Both the LR and RF results of TIRS show smaller RMSEs when compared to UGSG and MODIS LST data.Combined with 109,978 ETM+images,the spatial distributions of the annual average land surface emissivity,normalized difference vegetation index and LST with a spatial resolution of 30 m for the entire TP were obtained providing data with higher accuracy for the study of energy,water cycle and land-atmosphere interaction over the TP.Similar to LST,ETa is also an important component of the energy exchange process and hydrological cycle.The estimation of ETa is essential to the recognition of local climate change and the study of water and heat exchange cycle.How to obtain high-precision hourly ETa data is an important scientific issue.In this study,accurate hourly ETa data over the entire TP were proposed for the first time based on Fengyun4A geostationary satellite data,the Chinese Land Data Assimilation System(CLDASV2.0)and an RF model.Compared with the ETa results from the surface energy balance system(SEBS),maximum entropy production(MEP)and the European Centre for Medium-Range Weather Forecasts Reanalysis Fifth Generation(ERA5)with RMSE values of 1.997,2.182 and 1.555 mm/day,respectively,the RF model showed the best performance with an RMSE of 0.991 mm/day.According to the hourly evapotranspiration data proposed in this study,the annual ETa over the entire TP was 365.60 mm and the total water amount evapotranspired from the TP surface was approximately 9811.01 ×108 t/yr.Additionally,diurnal,monthly and seasonal ETa variations in different land cover types and climate zones over the TP and their contributions to the total ETa were clearly quantified.The ETa distribution presented rapid hourly variability in the daytime over the entire TP except in the Qaidam Basin,which was especially distinct in the southeastern TP and the central TP from 10:00 to 13:00 and from 15:00 to 18:00.ETa variations over these regions can exceed 4 mm/day.Diurnal,monthly and seasonal ETa variations in different land cover types and climate zones over the TP and their contributions to the total ETa were clearly quantified.The ETa in different regions has obvious diurnal variations during the daytime.Additionally,through the variable sensitivity of the machine learning model and correlation analysis,it is found that the net radiation is the main driving factor of ETa over the TP.