Topographic correction for remotely sensed imagery is an important preprocessing step in order to improve the retrieval accuracy of land surface spectral reflectance in mountainous area. Various kinds of topographic correction models have been proposed in the literature. Each model has its advantages and limitations. In consideration of the limitations of the topographic correction models in the literature, an improved Shepherd topographic correction model is proposed in this paper. Diffuse irradiance is an essential factor in the physically based topographic correction model. While in the Shepherd model (originally proposed by Shepherd et al. in 2003), accuracy of the method to compute the diffuse irradiance is relatively low; therefore, the accuracy of the land surface spectral reflectance retrieved with the Shepherd model is impacted. In order to improve the accuracy of diffuse irradiance, hence the accuracy of land surface spectral reflectance, a different method (named the Perez model), is used to obtain the diffuse irradiance with higher accuracy in the improved Shepherd model. Landsat 5 Thematic Mapper (TM) imagery acquired on July 12th 2006, over the mountainous areas in the north of Beijing city, was employed to retrieve land surface spectral reflectance with the improved Shepherd topographic correction model and 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) atmospheric radiative transfer model. Correction results were tested with three different methods. Testing result shows that the improved Shepherd topographic correction model can achieve a good correction result and is better than Shepherd and C topographic correction model. What is more, this improved model is physically based and can be applied to all kinds of optical satellite imagery.