Surface soil moisture (SSM) is a key state variable in controlling land surface energy balance and hydrological process. Based on the mechanism behind land surface temperature (LST)-vegetation index (VI) triangle space, an empirical relationship model has been proposed for SSM estimation with LST, NDVI, and surface albedo, and it has been applied in downscaling the coarse resolution microwave soil moisture product. In this paper, three soil moisture observation networks (REMEDHUS, MAQU, and MURRUMBIDGEE) were selected to evaluate the performance of this model at different climate and land cover conditions with in situ soil moisture measurements and Landsat satellite observations. According to the estimation results from different days for each network, it was found that the model was able to capture SSM variation with a satisfied accuracy [overall root-mean-squared error (RMSE) ranging from 0.025 to 0.055 m(3)/m(3)], and the R-2 can reach 0.9 on some individual days. However, the performance has high daily variability with some poor ones. The reason is partly attributed to the high sensitivity of the coefficients of the model to the variation degrees of the input LST, normalized difference vegetation index (NDVI), and SSM. Meanwhile, the spatial scale differences between the point measurement and satellite footprint observation are another important issue. To improve the model performance, a new relationship model was proposed by introducing the modified normalized difference water index, and the estimation results had a pronounced improvement (overall RMSE ranging from 0.021 to 0.049 m(3)/m(3)) compared with the previous model. The application effect of the proposed model showed that the model coefficient calibration accuracy greatly determined the uncertainty level of the estimation results.