Abstract: In the present study, a method named MECRBF was proposed based on mind evolutionary computation radial basis function neural network. Its ability to reveal spatial heterogeneity of soil organic matter was compared with radial basis function neural network (RBF-Near) based on spatial coordinates and neighbor information, and ordinary Kriging method with Wannian County, Jiangxi Province as study area. To establish and validate, 954 soil samples were collected and randomly divided into 2 groups, i.e. modeling points (763) and validation points (191). Spatial distribution prediction capacities and prediction map of these methods were compared. It was shown that the root mean square errors (RMSE), mean absolute errors (MAE) and mean relative errors (MRE) of MECRBF in validation points was 0.50 g·kg^-1, 0.39 g·kg^-1, and 1.40 percent smaller than those of RBF-Near (P<0.05), respectively, and was 2.59 g·kg^-1, 1.89 g·kg^-1, and 7.76 percent smaller than those of ordinary Kriging (P<0.05), respectively. The prediction map obtained by MECRBF was more consistent with the actual geographical information than the others. Moreover, MECRBF method reduced the prediction errors. The proposed MECRBF could provide guidance to predict soil nutrients at county scale.

Key words: soil organic matter, mind evolutionary computation, neural network, spatial heterogeneity