Hydroxyl radicals ( • OH) are prevalent in diverse ecosystems, including aquatic, atmospheric, and biological environments, and are crucial in the regulation of carbon and nitrogen cycles. However, there is still a lack of an easily applicable, efficient, and precise quantitative structure−activity relationship (QSAR) model for determining second-order reaction rate constants between • OH and pollutants. Herein, a quantitative QSAR model was constructed utilizing molecular images and a vision transformer in conjunction with density functional theory (DFT) to efficiently and precisely forecast second-order reaction rate constants involving hydroxyl radicals ( • OH) and pollutants. The model exhibits strong resilience and predictive precision achieved through transfer learning and fine-tuning, yielding test root-meansquare error values within the range of 0.2616−0.3239, surpassing the performance of the molecular image-convolutional neural network model. The reaction sites were accurately identified with a high level of precision, as evidenced by the F1-scores (>0.9775) and AUC-ROC values (>0.9665), along with validation using gradient-weighted class activation mapping and the Fukui function based on DFT. This research offers a cost-effective alternative to complex experimental methods and introduces a novel tool for environmental monitoring and risk assessment, highlighting its environmental significance and practical utility.