Modeling Long-term And High-resolution Land Surface Temperature And Precipitable Water Vapor Over The Tibetan Plateau

Due to the global warming,the land surface temperature(LST)and precipitable water vapor(PWV)changed a lot over the Tibetan Plateau.For monitoring the change in them,long-term,high-resolution and accurate LST and PWV data sets are crucial,but of lack.In this paper,models are built to produce long-term,high-resolution and accurate data sets.LST and relative humidity observed by Chinese meteorological stations over Tibetan Plateau(as ground observation)and LST and PWV from MODIS(as satellite observation)are used,since ground observations are available during long time and the satellite observations are characterized by high-resolution and high-accuracy.Before building models,ground observations are over-sampled to be balanced in space distribution,because imbalanced data sets weaken the model's prediction ability in minority area.And the technique about over-sampling is SMOTE.During the model-building session,both over-sampled ground and satellite observation during 2001-2010(10a)are used as the training data.And the Bayesian Linear Regression(BLR)is selected as the base regression model.Since the number of the over-sampled ground observation data dots is more than the original one,the concept "ensemble" is adopted.So LST ensemble Bayesian Linear Regression Model(LST-eBLR)and PWV ensemble Bayesian Linear Regression Model(PWV-eBLR)are trained.Then ground and satellite observation during 2011-2015(5a)are used to validate the models.Apply the models to the ground observation then model predictions are obtained.The satellite observations are set to be true value.Using "error" and"correlation" as the measures to evaluate the accuracy of model-prediction compared with the true value.The mean of errors of LST-eBLR prediction over the time from 2011 to 2015 is 0.21?,and 75 percent of the errors locates in[-0.84,1.26].And the mean of the correlation between LST-eBLR prediction and true value is 0.95,and 75 percent of the data dots' correlation locates in[0.90,0.99].As for the PWV-eBLR,the mean of errors is-0.59mm,and 75 percent of errors locates in[-0.77,0.60];the correlation between mean PWV-eBLR and true value is 0.9,and 75 percent of data dots'correlation locates in[0.82,0.95].So LST-eBLR and PWV-eBLR both possess high-accuracy prediction ability.Apply LST-eBLR and PWV-eBLR to ground observations during 1996-2000(5a)to produce high-resolution and accuracy LST and PWV.Along with the MODIS data sets during 2001 to 2015,long-term(1996-2015,20a),high-resolution and high accuracy LST and PWV data sets are obtained.ECMWF re-analysis data about LST and PWV over Tibetan Plateau are evaluated.LST-ECMWF and PWV-ECMWF both over-estimate the LST and PWV compared to LST-eBLR and PWV-eBLR.The mean of LST-ECMWF is 5.37?,twice more than the mean of LST-eBLR(which is 2.26 ?).And the long-term change rate of LST calculated by LST-ECMWF is 0.0026?/a,while LST-eBLR shows that the change rate is-0.0019?/a.As for the PWV,the mean of PWV over Tibetan Plateau is 12.42 mm,which is almost as twice as the mean of PWV-eBLR(6.82 mm).The change rate of PWV-ECMWF is 0.000mm/a,while the change rate of PWV-eBLR is-0.0045 mm/a.Despite the difference between the change rates of ECMWF re-analysis data sets and the model values,the LST and PWV changed little over the past 20 years.According to the rates calculated by eBLR models,the LST decreases about 0.0019?a year,and PWV decreases about 0.0045 mm a year.