Excessive application of chemical fertilizers triggers severe agricultural non-point pollution while securing crop yields. However, applying a small amount of field trial data to optimize fertilizer management strategies in complex environments is challenging, as yield and environmental risks must be balanced. Here, we develop a hybrid approach combining field trial data, transfer learning, and reinforcement learning (TL-RL) to adjust the fertilizer management algorithm dynamically and autonomously learn the optimal fertilizer application strategy through interaction with the environment. It leverages the effectiveness of field trial data, overcoming the disadvantage that reinforcement learning requires a large amount of data to train the model. We found that the optimal fertilization strategy involved a 13.9 % reduction in nitrogen fertilizer, 10.4 % reduction in phosphorus fertilizer, and 5.44 t/ha of hydrothermal humus charcoal, compared to the traditional fertilization approach. The optimal fertilization scheme resulted in a 7.02 % increase in rice yield, a 20.83 % reduction in total nitrogen concentration, and a 39.13 % reduction in total phosphorus concentration in the paddy surface water. We demonstrate that the TL-RL hybrid model provides a feasible modeling strategy for screening optimal fertilizer application schemes and offers a theoretical approach to achieving sustainable agricultural development by minimizing non-point source pollution while safeguarding crop yields.