Abstract: With the support of multi-scale segmentation of object-oriented classification method, features and characteristics variables were established and extracted by taking the QuickBird high-resolution remote sensing image as the main data source. High resolution remote sensing image classification model was built based on RF(Random Forest), the importance and sensitivity of characteristic variables to the model were analyzed. The research results showed that the optimal segmentation parameter in the study area was determined: segmentation scale parameter is 70, shape factor is 0.2, color factor is 0.8,and 32 characteristics such as spectrum, texture and shape variable information of eight landscape types were extracted; the overall accuracy of RF classification model which was built by choosing 5 000 trees and one node variable is 0.94, the Kappa coefficient was 0.93, OOB(Out of the Bag)data generalization error is 6.01%; through the analysis of the importance of variables, the importance value of spectral characteristics was obviously higher than that of shape features and texture features; based on mean decrease accuracy the overall accuracy is 0.94, the Kappa coefficient was 0.93 when 16 variables were used to built RF model; but they achieved the peak when 19 variables were used to built the model based on mean decrease Gini coefficient, the sensitivity analysis of the model which based on mean decrease accuracy was optimum than which based on mean decrease Gini coefficient. The feasibility of random forest model classification method should be further explored to provide scientific and rational reference for the high resolution remote sensing image classification.