Li Qi-Quan, Wang Chang-Quan, Zhang Wen-Jiang, Yu Yong, Li Bing, Yang Juan, Bai Gen-Chuan, Cai Yan
Sichuan Agricultural University, Chengdu, China.
Ying Yong Sheng Tai Xue Bao. 2013 Feb;24(2):459-66.
In this study, a radial basis function neural network model combined with ordinary kriging (RBFNN_OK) was adopted to predict the spatial distribution of soil nutrients (organic matter and total N) in a typical hilly region of Sichuan Basin, Southwest China, and the performance of this method was compared with that of ordinary kriging (OK) and regression kriging (RK). All the three methods produced the similar soil nutrient maps. However, as compared with those obtained by multiple linear regression model, the correlation coefficients between the measured values and the predicted values of soil organic matter and total N obtained by neural network model increased by 12. 3% and 16. 5% , respectively, suggesting that neural network model could more accurately capture the complicated relationships between soil nutrients and quantitative environmental factors. The error analyses of the prediction values of 469 validation points indicated that the mean absolute error (MAE) , mean relative error (MRE), and root mean squared error (RMSE) of RBFNN_OK were 6.9%, 7.4%, and 5. 1% (for soil organic matter), and 4.9%, 6.1% , and 4.6% (for soil total N) smaller than those of OK (P<0.01), and 2.4%, 2.6% , and 1.8% (for soil organic matter), and 2.1%, 2.8%, and 2.2% (for soil total N) smaller than those of RK, respectively (P<0.05).
本研究采用径向基函数神经网络模型结合普通克里金法（RBFNN_OK）预测中国西南部四川盆地典型丘陵区土壤养分（有机质和全氮）的空间分布，并将该方法的性能与普通克里金法（OK）和回归克里金法（RK）进行比较。三种方法生成的土壤养分图相似。然而，与多元线性回归模型得到的结果相比，神经网络模型得到的土壤有机质和全氮实测值与预测值之间的相关系数分别提高了12.3%和16.5%，表明神经网络模型能够更准确地捕捉土壤养分与定量环境因素之间的复杂关系。对469个验证点预测值的误差分析表明，RBFNN_OK的平均绝对误差（MAE）、平均相对误差（MRE）和均方根误差（RMSE）（土壤有机质）分别比OK小6.9%、7.4%和5.1%（P<0.01），（土壤全氮）分别比RK小4.9%、6.1%和4.6%（P<0.01），（土壤有机质）分别比RK小2.4%、2.6%和1.8%，（土壤全氮）分别比RK小2.1%、2.8%和2.2%（P<0.05）。
