基于辅助变量和回归径向基函数神经网络(R-RBFNN)的土壤有机质空间分布模拟

为快速准确地获取土壤有机质的空间分布状况,以江西省万年县齐埠镇为例,运用四方位搜索法、地统计学和遥感影像分析技术提取环境因子和邻近信息,构建基于环境因子和邻近信息的回归克里金法(RK)和回归径向基函数神经网络法(R-RBFNN),对齐埠镇耕地表层(0~20 cm)土壤有机质空间分布进行模拟,并与普通克里金法(OK)相比较。结果显示:齐埠镇耕地表层土壤有机质含量在17.30~53.58 g·kg^-1,平均值为35.03 g·kg^-1,变异系数为23.61%,呈中等变异性。半变异函数分析显示,土壤有机质的块金效应值为0.59,表现为中等空间相关性,自相关范围较大。利用62个采样点进行建模、16个采样点进行独立验证,误差分析表明,应用环境因子和邻近信息作为辅助变量的RK和R-RBFNN预测结果的均方根误差、平均绝对误差、平均相对误均差较OK降低,测试集中的相对提高度分别为66.67%和71.79%,显示出较高精度。但R-RBFNN无须计算半方差函数,使用简单,因此更具优势。

Prediction of soil organic matter distribution based on auxiliary variables and regression-radial basis function neural network (R-RBFNN) model

Accurate spatial distribution information about soil organic matter (SOM) is critical for farmland use and soil environmental protection. In order to find the best interpolation method of SOM in Qibu Town in Wannian County, Jiangxi Province, a regression-radial basis function neural network (R-RBFNN) model was proposed based on environmental factors and neighbor information, regression Kriging (RK), based on environmental factors and neighbor information, and ordinary Kriging (OK) were also dopted to predict SOM distribution. Environmental factors were extracted from digital terrain and remote sensing image, and the four-direction search method was applied to get the neighbor information. To establish and validate this method, 78 soil samples were collected and randomly divided into two groups, as modeling points (62) and validation points (16). Results showed that, SOM content ranged from 17.30 to 53.58 g·kg^-1, with an average of 35.03 g·kg^-1, indicating a moderate variability. The nugget/sill ratio was 0.59, indicating a moderate spatial dependence for SOM. The prediction map obtained by RK and R-RBFNN was similar and more consistent with the true geographical information than OK. Moreover, compared to OK model with the validation points, RK model and R-RBFNN reduced the prediction errors, as the root mean square errors (RMSE), the mean absolute errors (MAE) and the mean relative errors (MRE) of RK were all reduced to those of OK, and the relative improvement was 66.67% and 71.79%, respectively. Therefore, both RK and R-RBFNN significantly improved the interpolation accuracy of SOM distribution due to the consideration of the environmental factors and neighbor information. In addition, R-RBFNN did not require calculation of semi-variogram, and thus exhibited better application potential.