A neural network approach to simple prediction of soil nitrification potential: A case study in Japanese temperate forests

The nitrogen status of forest ecosystems can be represented by the net nitrification potential (NNP) of the forest soils. Prediction of NNP using a small number of soil properties is a practically useful tool for forest management planning. Artificial neural networks (ANN) have recently become popular tools in forest modeling because they eliminate certain difficulties in handling forest data, such as the nonlinear relationships and non-normality. This study aimed to develop an ANN model to predict NNP that required a few soil properties as possible for input data. The ANN model was fitted to field data using the ridge-stabilized Gauss–Newton method, with a subset of methods to prevent excessively high weights that are likely to cause over-fitting. We collected surface mineral soil samples from 56 locations in temperate forest ecosystems of central Japan. We measured NNPs on a per area basis (Mg N km−2) using aerobic laboratory incubation at 30 °C for 4 weeks. The ANN-based model using data on only two soil properties (the C:N ratio and the maximum water-holding capacity) provided the best prediction of NNP. The ANN-based model’s success results from its incorporation of (1) the nonlinear relationship between the C:N ratio and NNP and (2) the hierarchical control of NNP, which is governed primarily by the C:N ratio and secondarily by soil moisture conditions. The simplicity of the model greatly enhances its practical value in forest management planning.