To investigate the regional variations in site effects and establish empirical site amplification models, a large number of seismic ground motion recordings collected by the National Strong-Motion Observation Network System of China from 2007 to 2020 were utilized to separate the site effects at 211 strong-motion stations mainly distributed in the western part of the Southern Tianshan, Longmenshan fault zone, and Yunnan based on the generalized spectral inversion method. The empirical models of site effects associated with either the site category or V_S30 were established. Furthermore, the regional dependence of site effects was systematically investigated, along with the correlation between site amplification characteristics and the code-specified site classification indicators (overburden thickness H and equivalent shear-wave velocity V_Se). The results demonstrate that the empirical amplification models for Site Classes Ⅰ, Ⅱ, and Ⅲ effectively capture the systematic differences in site effects across the entire frequency spectrum. Significant regional variations in site amplification are observed within the same site classification category, with particularly pronounced differences between Class III sites in the western Southern Tianshan and Yunnan regions. The correlation between the site classification indicators (H and V_Se) and actual site amplification characteristics is found to be statistically insignificant, with potential regional discrepancies noted in this relationship. Meanwhile, the V_S30-based empirical models reliably represent the influence of local site conditions on ground motion amplification. These results provide a scientific basis for developing regionalized ground motion prediction models, improving the accuracy of seismic risk assessments and guiding the practical application of regional site amplification effects. Furthermore, the results offer valuable support for exploring more effective site classification indicators and establishing regionally differentiated site classification standards.