Abstract: Imbalanced nutrient supply during the growth period of pak choi often resulted in low fertilizer utilization efficiency and limited the accuracy of fertilizer application rate prediction. Conventional fertilization strategies relied heavily on empirical decision-making and offline soil analysis, which could not meet the requirements of precision agriculture under dynamic field conditions. To address these challenges, this study aimed to develop an intelligent variable-rate fertilization system capable of integrating real-time soil nutrient sensing with accurate fertilizer demand prediction. In particular, the objective was to improve the predictive performance of fertilizer application models by combining machine learning techniques with an improved swarm intelligence optimization strategy, thereby providing technical support for rapid, precise, and efficient fertilization management of pak choi. A variable-rate fertilization machine based on real-time soil nutrient detection was constructed. The system was designed to accomplish online soil parameter acquisition, fertilizer demand prediction, and fertilizer execution within a single operation cycle. Historical growth environment data and soil nutrient information of pak choi were collected and preprocessed to establish a fertilizer recommendation dataset, at Anhui Agricultural University in Hefei, China (117°25′E, 31°87′N). A back propagation neural network was employed as the core predictive model, and its network parameters were optimized using a hybrid Sooty Tern Optimization Algorithm. The proposed hybrid optimization approach was introduced to enhance convergence stability and avoid premature local optima. The developed hybrid-optimized neural network model was trained and validated using experimental data, and its performance was systematically compared with a standard back propagation neural network, a whale-optimized back propagation neural network, and a sooty tern-optimized back propagation neural network under identical conditions. Furthermore, systematic field experiments were conducted to assess fertilizer discharge stability and operational performance of the developed device at three representative operating speeds, namely low, medium, and high travel speeds, to verify the applicability of the proposed strategy under practical working conditions. The experimental results demonstrated that the proposed hybrid Sooty Tern Optimization Algorithm–optimized back propagation neural network achieved significantly improved prediction accuracy for fertilizer application rates. The predicted fertilizer amounts exhibited a high degree of consistency with the actual application values across different growth stages of pak choi. The coefficient of determination of the proposed model reached 0.970, which was markedly higher than those of the sooty tern–optimized model, the whale-optimized model, and the conventional neural network model, whose coefficients of determination were 0.867, 0.814, and 0.655, respectively. Furthermore, the maximum absolute percentage error of the proposed model was limited to 9.89%, which was substantially lower than the corresponding errors of 17.17%, 18.15%, and 24.19% observed in the comparison models. These results indicated that the hybrid optimization strategy effectively enhanced both global search capability and local refinement of the neural network parameters. The integration of real-time soil sensing with the optimized prediction model enabled more accurate fertilizer demand estimation and improved adaptability to complex field environments. The hybrid Sooty Tern Optimization Algorithm–based back propagation neural network proposed in this study significantly improved the accuracy and robustness of fertilizer application rate prediction for pak choi. The developed variable-rate fertilization system demonstrated strong potential for synchronizing soil information acquisition, fertilizer demand modeling, and fertilization execution in real time. Field performance tests further confirmed the feasibility of the proposed fertilization strategy. At the lowest operating speed, the comprehensive fertilizer discharge accuracy reached 97.5%, indicating excellent stability and consistency. When operating at a medium speed, the overall accuracy remained high at 95.1%. Although an increase in travel speed resulted in a gradual decline in discharge precision, the comprehensive accuracy still maintained a satisfactory level of 91.0% under the highest tested speed. The results confirmed that the proposed fertilization strategy model could effectively support precise nutrient management and enhance fertilizer use efficiency. This study provided a feasible and reliable technical reference for the implementation of intelligent variable-rate fertilization systems in leafy vegetable production.