Spatial prediction of soil nutrient in a hilly area using artificial neural network model combined with kriging

• Published in: Archiv für Acker- und Pflanzenbau und Bodenkunde 2016-11, Vol.62 (11), p.1541-1553
• Main Authors: Li, Qi-Quan, Zhang, Xin, Wang, Chang-Quan, Li, Bing, Gao, Xue-Song, Yuan, Da-Gang, Luo, You-Lin
• Format: Article
• Language: English
• Subjects: auxiliary environmental factors; Environmental factors; kriging; Mean square errors; Neural networks; Radial basis function neural network; soil nutrient; Soil nutrients; Soil properties; Soil sciences; Spatial distribution; spatial prediction
• ISSN: 0365-0340; 1476-3567
• DOI: 10.1080/03650340.2016.1154543

It is widely recognized that using correlated environmental factors as auxiliary variables can improve the prediction accuracy of soil properties. In this study, a radial basis function neural network (RBFNN) model combined with ordinary kriging (OK) was proposed to predict spatial distribution of four soil nutrients based on the same framework used by regression kriging (RK). In RBFNN_OK, RBFNN model was used to explain the spatial variability caused by the selected auxiliary factors, while OK was used to express the spatial autocorrelation in RBFNN prediction residuals. The results showed that both RBFNN_OK and RK presented prediction maps with more details. However, RK does not always obtain mean errors (MEs) which were closer to 0 and lower root mean square errors (RMSEs) and mean relative errors (MREs) than OK. Conversely, MREs of RBFNN_OK were much closer to 0 and its RMSEs and MREs were relatively lower than OK and RK. The results suggest that RBFNN_OK is a more unbiased method with more stable prediction performance as well as improvement of prediction accuracy, which also indicates that artificial neural network model is more appropriate than regression model to capture relationships between soil variables and environmental factors. Therefore, RBFNN_OK may provide a useful framework for predicting soil properties.