With the forest inventory data and relevant remote sensing data, two approaches of variable selection, i.e. bootstrap method and variable importance of projection (VIP) criterion, were compared in their impacts on prediction accuracy of estimating model.Subsequently, the forest canopy closure density of the study area, Tahe County, was estimated by a simplified PLSR(partial least square regression) model,which has a better prediction accuracy. Here, we defined a model including all variables as full model, a model only including selected variables by Bootstrap method as Bootstrap model, and a model only including selected variables by VIP criterion as VIP model. The results showed that all those variables selected by VIP criterion passed through non-parameters Boostrap test (α=0.05). The VIP model had no lower accuracy than full model and Bootstrap model, and the relative root mean square error of the former was 99.2% and 99.6% of those of the latters, respectively. In addition, the VIP model improved the estimation accuracy at plot or pixel level.