The traditional spectral ratio method for seismic site classification often involves subjectivity and uncertainty. By integrating machine learning algorithms and leveraging large datasets along with feature extraction techniques,the objectivity and accuracy of the classification process can be enhanced,providing a more reliable foundation for seismic engineering design and hazard assessment. In this study,strong-motion waveform data from historical earthquakes in the Sichuan-Yunnan region were used to generate spectral ratio curves by calculating the horizontal and vertical velocity response spectrum values for various site categories. The characteristics of these curves,including predominant period,peak,and overall values,were analyzed. By incorporating deep learning algorithms,a dataset of spectral ratio curves from strong-motion stations in the Sichuan-Yunnan region was compiled and used as input for the VGG16 convolutional neural network(CNN)model. This model was trained to predict the probability distribution of three site categories based on NEHRP standards. The accuracy and generalization ability of the model were further enhanced through feature extraction and cross-validation techniques. The resulting site classification model demonstrated strong performance in validation,offering significant reference value for improving and applying the spectral ratio method in seismic site classification.