Transpiration, an essential component of surface evapotranspiration, is particularly important in the research of surface evapotranspiration in arid areas. The paper explores the spectral information of the arid vegetal evapotranspiration from a semi-empirical perspective by the measured data and the up-scaling method. The paper inverted the transpiration of Haloxylon ammodendronat at the canopy, pixel and regional scales in the southern edge of the Gurbantunggut desert in Xinjiang, China. The results are as follows: At the canopy scale, the optimal exponential model of the sap flow based on the hyperspectrum is Y = 3.65× SR(1580,1600) + 0.76, R2 = 0.72. At the pixel scale, there was a good linear relationship between the sap flow and the SR index, with a linear relationship of Y = 0.0787 X - 0.0724, R2 = 0.604. At the regional scale, based on the optimal exponential model and the EO-1 Hyperion remote sensing data, the transpiration of the study area was inverted. Comparing the results of the SEBAL and SEBS models, the errors of the simulation results were 12.66% and 11.68%. The paper made full use of the knowledge flow at different scales, bridging the scale difference in canopy and remote sensing images to avoid the information bottleneck in the up-scaling. However, there is much limit in the data acquirement, the endmembers determine, the temporal-spatial up-scaling, and the accuracy assessment to be improved in the future studies.