Pedogeography and Hydrology
LI Qiquan, WANG Changquan, YUE Tianxiang, LI Bing, ZHANG Xin, GAO Xuesong, ZHANG Yi, YUAN Dagang
Soil organic matter (SOM) is one of the most important indicators of soil quality. Accurate spatial information about SOM is critical for sustainable soil utilization and management and environmental protection. Spatially correlated auxiliary information was widely used to improve spatial prediction accuracy. However, the qualitative variables such as soil type, land use type are not being used often as auxiliary variables. In this paper we proposed a spatial prediction method (ST+RBFNN) based on radial basis functional neural network model, using both qualitative and quantitative variables as auxiliary information, to predict the spatial distribution of soil organic matter in Santai County in Sichuan Province, located in the hilly region of mid Sichuan Basin. To establish and validate this method, 2346 soil samples were collected and randomly divided into two groups, as modeling points (1877) and validation points (469). With the modeling points, a radial basis function neural network model was trained using the average content of SOM of each soil genus, topographical factors and vegetation index as auxiliary information to predict the spatial distribution of SOM content within each soil genus. Results showed that, the SOM content ranged from 4.20 to 47.60 g kg-1, with an average value of 17.97 g kg-1, a moderate variability. The nugget/sill ratio was 0.742, indicating a weak spatial dependence for SOM. Elevation and slope showed significantly negative correlation with SOM content while topographic wetness index and vegetation index showed significantly positive correlation with SOM. Analysis of variance indicated that there were significant differences in average content of SOM among the different soil types (P<0.01), suggesting that soil types also had significant impact on the spatial distribution of SOM, and soil genus types were better predictors than soil groups. Slope, topographic wetness index and vegetation index showed significant correction with the residuals of average content of SOM (computed by subtracting the average SOM content of the relative soil genus from the original value of each soil sample), indicating that the above three quantitative factors further resulted in the spatial variation of SOM besides soil types. The prediction map obtained by the proposed method was more consistent with the true geographical information than ordinary Kriging (OK), regression Kriging (RK) and neural network combined with ordinary Kriging (RBFNN+OK). Moreover, ST+RBFNN method significantly reduced the prediction errors. Compared to OK, RK and RBFNN+OK, the mean absolute error (MAE) of ST+ RBFNN method was reduced by 31.76%, 28.45% and 26.68%, the mean relative error (MRE) was reduced by 35.90%, 32.55% and 30.75%, and the root mean squared error (RMSE) was reduced by 22.60%, 19.88% and 18.43%. Moreover, this method also showed better capability of predicting the extremum of the validation data. The prediction errors were reduced by 6.88% to 43.70% than the other three methods in predicting the extremum of the validation points (10% of normal distribution of the data). This result suggested that it is helpful for improving the prediction accuracy to employ both qualitative and quantitative variables as auxiliary information in spatial prediction of soil properties, and this proposed method provides a useful research idea for digital soil mapping.