5. The evaluation of Model

5.1 The check with independent sample
It will randomly take out some samples for the check model before them being sent into the model for acquiring regression number of independent samples and the error with the input of coefficients for each mode. The result shows the model is stable as the mean square errors at both table 3 and 4 are quite close.

5.2 Coefficient rule of composite correlation
The coefficient of composite correlation is a relative index for showing the regression result. It indicates that the bigger the coefficient of composite correlation, the closer the statistical relation between independent variable and dependent variable. In the table 3, the coefficients of composite correlation are all bigger than 70% except the one in cloud day of July for zone 1 is just 67%. In the study, the coefficient of composite correlation for summer is higher than the one of winter and in clear day is bigger than cloud day.

5.3 Mean square error
The mean square error is an index for directly showing the accuracy of simulated data. At the table 3, the mean square errors in july are all below 2.5°C except the one of cloud day for zone 1. The mean square errors for clear day of January are all smaller than 2.8°C and the mean square errors of cloud day of January are below 3.6°C, which anyhow is still much smaller that the ones derived from samples. In our study, it indicates that the mean square error of summer is better than winter and for clear day is better than cloud day.

5.4 Test of the composite correlation coefficient - statistic F.
The regression result is tested by the statistic F with the variance analysis of line regression.

in the equation , m is the independent variable, n is number of sample and R is the composite correlation coefficient. The regression result is good as each F value is bigger than Fa value.

As the representativeness of conventional observations stations is not so good due to the inhomogeneous Spatial sampling, it is necessary to remove the bad ones with the quality control process. But it also cause a worse representativeness of the plateau as many samples are filtered out in the quality control process. For solving the problem, all data that can be matched with satellite data are input into the model. The resulted of the model indicates that it is pretty ideal.

6. Conclusion
The model based on NOAA/TOVS data is successfully for the prediction of daily maximum shelter temperature of Tibetan Plateau. The result from the model has more representativeness and homogeneity than the conventional data and its accuracy is also better than some existing methods. It has no anormal data (cloud contaminated data) in the matched data that used for the regression analysis of the model, which means the cloud detection technique accepted in the model is good. As the microwave data plays a main rule in the model, it indicates that the shelter temperature of all conventional observation time levels over Tibetan Plateau can be retrieved from satellite data by the regression method.

References

• Liu Ruiyun, Fan Tianxi, Retrieval of Maximum Land Surface Temperature with NOAA Satellite Data, Journal of Najing Institute of Meteorology, P106- 110, Vol.19, No.1, 1996.
• P.K. Rao, S.J. Holmes, Weather Satellites: System Data, and Environmental Application, American Meteorological Society, 1990.
• Liu Ruiyun, Luo Jingning, Guo Lujun, Monitoring Snow Cover with TOVS Data, Quarterly Journal of Applied Meteorology, p88-93, Vol. 10, No.1, 1999.

Detector	channel	central frequency	peak energy	name
			contribution height (HPa)
MSU	1	50.31	surface	TB1
	2	53.73	700	TB2
	3	54.96	300	TB3
	4	57.95	90	TB4
HIRS/2	8	898(cm-1)	surface	TB8
	9	1028	25	TB9
	11	1364	700	TB11
	14	2213	950	TB14
	long wave radiation derived from			OLR
	multi channel calculation

Table 1. Wavelength of related channels, peak energy contribution layer and name of brightness temperature.