Abstract: In order to solve the problem of inaccurate results in the traditional regression model for optimization of soybean planting density and fertilizer application, a BP-Linear Constrained Optimization（BP-LCO） method based on BP neural network was proposed in this study. We used Heihe 43 as experimental material to carry out four-factor and five-level orthogonal rotation test. The experimental factors were planting density of soybean, application amount of N, P₂O₅ and K₂O, and the evaluation index was yield of soybean, BP-LCO algorithm was used to construct a fitting model for the relationship between planting density, fertilizer application and yield, and we carried out global optimization and validation experiments. The results of the model analysis showed that the optimal planting density was 36.67×10~4 plants·ha^-1, N application rate was 77.98 kg·ha^-1, P₂O₅ application rate was 93.79 kg·ha^-1, K₂O application rate was 24.34 kg·ha^-1, and the corresponding soybean yield was 3 679.56 kg·ha^-1. The verification test showed that the actual yield of soybean was 3 702.29 kg·ha^-1 under the optimal ratio, and the relative error between the actual yield and the theoretical yield was 0.62%, which proved that the method had accurate optimization results and was an effective optimization method.