Artificial neural network and time series models for predicting soil salt and water content. (Special Issue: Salinity management in China.)

摘要

Volumetric water content of a silt loam soil (fluvo-aquic soil) in North China Plain was measured in situ by L-520 neutron probe (made in China) at three depths in the crop rootzone during a lysimeter experiment from 2001 to 2006. The electrical conductivity of the soil water (EC_sw) was measured by salinity sensors buried in the soil during the same period at 10, 20, 45 and 70 cm depth below soil surface. These data were used to test two mathematical procedures to predict water content and soil water salinity at depths of interest: all the available data were divided into training and testing datasets, then back propagation neural networks (BPNNs) were optimized by sensitivity analysis to minimizing the performance error, and then were finally used to predict soil water and EC_sw. In order to meet with the prerequisite of autoregressive integrated moving average (ARIMA) model, firstly, original soil water content and EC_sw time series were likewise transformed to obtain stationary series. Subsequently, the transformed time series were used to conduct analysis in frequency domain to obtain the parameters of the ARIMA models for the purposes of using the ARIMA model to predict soil water content and EC_sw. Based on the statistical parameters used to assess model performance, the BPNN model performed better in predicting the average water content than the ARIMA model: coefficient of determination (R2)=0.8987, sum of squares error (SSE)=0.000009, and mean absolute error (MAE)=0.000967 for BPNN as compared to R2=0.8867, SSE=0.000043, MAE=0.002211 for ARIMA. The BPNN model also performed better than the ARIMA model in predicting average EC_sw of soil profile. However, the ARIMA model performed better than the BPNN models in predicting soil water content at the depth of 20 cm and EC_sw at the depth of 10 cm below soil surface. Overall, the model developed by BPNN network showed its advantage of less parameter input, nonlinearity, simple model structure and good prediction of soil EC_sw and water content, and it gave an alternative method in forecasting soil water and salt dynamics to those based on deterministic models based on Richards' equation and Darcy's law provided climatic, cropping patterns, salinity of the irrigation water and irrigation management are very similar from one year to the next.Digital Object Identifier http://dx.doi.org/10.1016/j.agwat.2010.02.011