To address issues in traditional tire noise prediction methods such as simplified assumptions in physical modeling, insufficient generalization ability of data-driven models with small samples, and lack of multi-band noise collaborative analysis, this study proposes a multimodal tire noise prediction method that integrates transfer learning with physical mechanisms. By constructing a transfer learning network to extract geometric topological features of the tread pattern combined with a channel attention mechanism for dynamic weighting, an innovative physically-constrained hybrid loss function and multi-objective regression architecture are designed. Based on 126 sets of semi-anechoic chamber experimental data, a grouped cross-validation strategy is used to validate model performance. The results show that the combination of ResNet101 and LightGBM achieves the best performance in cross-modal feature fusion, with an average absolute prediction error of 0.66 dB, a coefficient of determination R2 of 0.97, and 96.96% of samples having prediction errors below 2 dB. The physical fundamental frequency constraint reduces model error by 15%, effectively balancing the advantages of data-driven and mechanistic models.