Objective  The aims were to predict and make grid soil organic matter (SOM) distribution map by using natural soil forming variables, in order to develop the tropical digital precision agriculture.

  Method  Based on the data of soil testing formula samples in Jinghong City, Yunnan Province in 2006, BP neural network (BPNN), BP neural network model based on strong classification algorithm (BPNN-Ada), BP neural network based on particle swarm optimization (PSO-BPNN), BP neural network based on genetic algorithm (GA-BPNN) and multiple linear regression (MLR) were used to predict and spatially map the SOM content.

  Result  ① The X and Y coordinate values of soil samples could improve the accuracy of the algorithm effectively, and express the cooperative relationship between environmental factors and the spatial distribution of SOM. ② The spatial distribution of SOM predicted by the four neural network algorithms was basically similar, showing a trend of high in the South and low in the north. ③ The fitting degree of the four neural network models in the modeling set in the study area from high to low followed: BPNN-Ada> GA-BPNN > PSO-BPNN > BPNN. In the modeling set, the fitting accuracy of pao-bpnn and GA-BPNN was consistent with that of BPNN, and the highest fitting accuracy R² of BPNN ADA was 0.98. In the verification set, the prediction ability from high to low followed: BPNN-Ada > GA-BPNN > PSO-BPNN > BPNN. BPNN-Ada had the highest prediction accuracy and algorithm stability, and the order was: RMSE = 4.47, MAE = 3.3, MRE = 0.05, R² = 0.98.

  Conclusion  The addition of geographical coordinates to the neural network modeling of SOM in Yunnan can effectively improve the accuracy of the model, and the neural network optimization algorithm based on learning rules is better than the neural network algorithm that optimizes initial weights and thresholds and the traditional BPNN algorithm.