Abstract: Imbalanced nutrient supply during the growth period of Pak choi can cause low efficiency of the fertilizer application. It is often required to predict the accuracy of the fertilizer application rate. However, the conventional fertilization strategies cannot fully meet the requirements of precision agriculture under dynamic field conditions. Particularly, they have relied heavily on empirical decision-making and offline soil analysis. In this study, an intelligent variable-rate fertilization was developed to integrate the real-time soil nutrient sensing with the accurate prediction of the fertilizer demand. The performance of the fertilizer application was improved to combine the machine learning with an improved swarm intelligence optimization. Thereby, the technical support was provided for the rapid, precise, and efficient fertilization of the Pak choi. A variable-rate fertilization machine was constructed to detect the soil nutrients. The soil parameters were online acquired to predict the fertilizer demand. The fertilizer was applied in a single operation cycle. Historical growth environment and soil nutrient data of the Pak choi were collected at Anhui Agricultural University in Hefei Province, China (117°25′E, 31°87′N). The fertilizer recommendation dataset was obtained after preprocessing. A back-propagation neural network was employed as the core predictive model. Its network parameters were then optimized using a hybrid Sooty Tern Optimization. The hybrid optimization approach was introduced to enhance the convergence stability and avoid the premature local optima. After that, the hybrid neural network model, after optimization, was trained and then validated using experimental data. The 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. Furthermore, the systematic field experiments were conducted to assess the fertilizer discharge stability and operational performance of the developed device at the low, medium, and high travel speeds, in order to verify the applicability under practical working conditions. The experimental results demonstrated that the hybrid Sooty Tern Optimization Algorithm achieved high prediction accuracy for the fertilizer application rates. The fertilizer amounts exhibited a high degree of consistency with the actual application over the different growth stages of the Pak choi. The coefficient of determination reached 0.970, which was markedly higher than those of the sooty tern model (0.867), the whale model (0.814), and the conventional neural network model (0.655), respectively. Furthermore, the maximum absolute percentage error was limited to 9.89%, which was substantially lower than those of 17.17%, 18.15%, and 24.19%, respectively, compared with the rest of the models. The hybrid optimization effectively enhanced both global search and local refinement of the neural network parameters. The soil sensing was real-time integrated with the prediction model. More accurate fertilizer demand was estimated under complex field environments. The high accuracy and robustness of the fertilizer application rate were obtained for the Pak choi. The variable-rate fertilization can be expected to synchronously acquire the soil information, fertilizer demand prediction, and execution in real time. Field performance tests further confirmed the feasibility of the fertilization strategy. Furthermore, the fertilizer discharge accuracy reached 97.5 % at the lowest operating speed, indicating the excellent stability and consistency. The overall accuracy remained high at 95.1 % when operating at a medium speed. Since the increasing travel speed caused the gradual decline in the discharge precision, the high accuracy was still maintained at an optimal level of 91.0 % under the highest tested speed. The strategy model effectively supported the precise nutrient management for the fertilizer use efficiency. This finding can provide a feasible and reliable technical reference to implement the intelligent variable-rate fertilization in leafy vegetable production.