Building damage after seismic disasters is one of the most vital factors threatening people's lives. The seismic vulnerability of buildings over large scales at the regional or country level is a key parameter for the mitigation of seismic disaster risk and rapid assessment of casualties after seismic events. To acquire the seismic vulnerability of buildings over large areas, a machine learning method based on mid-resolution satellite optical images is proposed. Taking field-investigated building vulnerability at the satellite pixel scale as a reference, the 15 most correlated features are calculated based on VIIRS nighttime light, MODIS vegetation index and surface reflectance, and texture data from Landsat-8 OLI surface reflectance products. Taking Yancheng, Jiangsu Province, China, as the study area, where 401 pixel-level seismic vulnerabilities (PLSVs) of the building environment are acquired based on field investigations, support vector regression (SVR) and random forest (RF) models are proposed using the 15 features calculated from satellite optical images. The results show that the proposed method can be used to estimate the PLSV with a root mean square error of approximately 0.1, with the PLSV normalized between 0 and 1. The machine learning model proposed in this study has a better accuracy for PLSV estimation than spatial interpolation.