Air Temperature Estimation over the Third Pole Using MODIS LST

Abstract:

The Third Pole is centered on the Tibetan Plateau (TP), which is the highest large plateau around the world with extremely complex terrain and climate conditions, resulting in very scarce meteorological stations especially in the vast west region. For these unobserved areas, the remotely sensed land surface temperature (LST) can greatly contribute to air temperature estimation. In our research we utilized the MODIS LST production from both TERRA and AQUA to estimate daily mean air temperature over the TP using multiple statistical models. Other variables used in the models include longitudes, latitudes, Julian day, solar zenith, NDVI and elevation. To select a relatively optimal model, we chose six popular and representative statistical models as candidate models including the multiple linear regression (MLR), the partial least squares regression (PLS), back propagate neural network (BPNN), support vector regression (SVR), random forests (RF) and Cubist regression (CR). The performances of the six models were compared for each possible combination of LSTs at four satellite pass times and two quality situations. Eventually a ranking table consisting of optimal models for each LST combination and quality situation was built up based on the validation results. By this means, the final production is generated providing daily mean air temperature with the least cloud blockage and acceptable accuracy. The average RMSEs of cross validation are mostly around 2℃. Stratified validations were also performed to test the expansibility to unobserved and high-altitude areas of the final models selected.