1：Experimental Design and Method Selection:

Models were constructed following the SVM approach. The retrieving scheme and the calculation process are described in Figs. 1 and 2, respectively. The system contains 16 CPUs and four GPUs (C2075) each with 448 cores. All data vectors were established, sorted, and standardized according to our design. A small fraction of the total data points (1/6 to 1/10) were selected as training datasets, and the remaining served as validation datasets. The SVM programmer (SVM Light V

2：01, regression) was used to teach the training datasets, i.e., to develop models, and the models were then tested with the validation datasets. An iterative procedure allowed for model parameter to be adjusted based on repetition of both training and testing segments to improve accuracy of the ?nal model. The model depicts results in the form of a linear scatter plot. Sample Selection and Data Sources:

Parameters that describe the characteristics of atmosphere and Earth surface are potential in?uential variables. MODIS data provide accurate information about our environment. They are used to understand dynamics and processes occurring on land, in oceans, and in the lower atmosphere. National Satellite Meteorology Center of China provides Earth imagery data (hourly scenes) acquired by FY satellites. In the current study, direct MODIS data (from AQUA and TERRA) obtained from Guangzhou Receiving Station were selected to be modeled. In addition, DEM data obtained from Chinese Research Academy of Environmental Sciences were used to create slope steepness, aspect, sun shadow, solar elevation and its sine (the product of solar elevation and the sine of solar elevation angle), horizontal solar orientation, slope solar elevation and its sine, hourly slope solar radiation energy (per unit area at the set time point assuming that there is absolutely no atmospheric layer), and preaccumulative solar radiation energy (the sum of positive solar radiation energy per unit area prior to the set time point). Finally, based on the astronomical and calendrical rules, accumulative solar radiation energy (per unit area of that day), and sunrise, sunset, and day time length were calculated and selected as in?uential variables.

3：List of Experimental Equipment and Materials:

AMAX high-performance supercomputer workstation, Windows 7 operating system, SVM light V

4：01 for model construction, ArcGIS 3, Arc Engine 3 and GDAL9 for GIS spatial data analysis and GIS data applications, VS2008C++/C#, MS Excel 2007 for input data collection and management, and applications of parallel computing. Experimental Procedures and Operational Workflow:

The study area was divided into 37 square subset areas (sBlk). Each sBlk is further divided into 1202 × 1202 pixels (sBlkSamples) to obtain the predicted temperatures and to construct the Ta image.

5：Data Analysis Methods:

The results are closely distributed on the 1:1 line with statistically signi?cance (R2 = 0.989) indicating low prediction noise. The errors were apparently not affected by temperature values.