Abstract: Abstract: Agricultural products processing is the key node in food supply chain quality control and products recall is the last shelter to protect the safety of consumers. The potential recall can be reduced by optimization of batch dispersion in the production planning stage of agricultural product processing based on satisfying the processing process and the processing technology. In this paper, the recall optimization in agricultural products processing node was researched. For the recall optimization in agricultural products processing, a four-level batch dispersion model was given by reference the food batch dispersion model. The model consists of four levels, respectively raw materials, components, semi-finished products and finished products. The model has there operations, respectively disassembling, assembling and packaging. In each level, composed several batches belong to different types and batches have ID, type and size three properties. The optimization model of four-level model is given, and its computational complexity was analyzed. According to the batch dispersion model is a NP hard problem, difficulty to resolve, particle swarm optimization was referred to solve it. In the early stage of PSO evolution, the fast convergence can easily cause premature, but in the later stage, slow convergence can easily cause oscillations. A modified PSO based on piecewise threshold particle replacement strategy was proposed to solve this problem. The formal description and diagrammatic of particle replacement strategy based on piecewise threshold were given and the setting of the threshold was discussed. Because the optimization model has more constraints and most of them are linear constraints, the penalty functions were introduced to solve the linear equality and inequality constraints in the recall optimization model. Relevant example was used to verify MPSO, and performance was compared with similar intelligent optimization algorithms. Numerical simulation and performance comparison show that the algorithm is efficient and can reduce the potential recall sizes about 30% on the base of reducing computing overhead about 90%, which is applicable to the recall optimization of agricultural products processing.