Surface evapotranspiration (ET) is one of the key surface processes. Reliable estimation of regional ET solely from satellite data remains a challenge. This study applies recently proposed nonparametric (NP) approach to retrieve surface ET, in terms of latent heat flux (LE), over a semiarid region. The involved input parameters are surface net radiation, land surface temperature, near-surface air temperature, and soil heat flux, all of which are retrievals or products of the Moderate-Resolution Imaging Spectroradiometer (MODIS). Field observations are used as ground references, which were obtained from six eddy covariance (EC) sites with different land covers including desert, Gobi, village, orchard, vegetable field, and wetland. Our results show that the accuracy of LE retrievals varies with EC sites with a determination of coefficient from 0.02 to 0.76, a bias from −221.56 W/m2 to 143.77 W/m2, a relative error from 8.82% to 48.35%, and a root mean square error from 67.97 W/m2 to 239.55 W/m2. The error mainly resulted from the uncertainties from MODIS products or the retrieval of net radiation and soil heat flux in nonvegetated region. It highlights the importance of accurate retrieval of the input parameters from satellite data, which are the ongoing tasks of remote sensing community.