Application of BP neural network in forecasting soil temperature time-series

Soil temperature is an important parameter in agro-meteorological observations and research on variations in soil temperature time series provides the theoretical basis for soil solute movement and proper distribution of reference crops. Owing to the complicated characteristics and strong nonlinearity of surface soil temperature time-series in saline coastal regions, a long-term data on soil temperature was analyzed and soil temperature time-series predicted by integrated Back-Propagation (BP) neural network method and expansion of multi-dimensional time-series, using experimental data from a typical coastal region in North Jiangsu Province. The results show that BP neural network model with 7 neurons of input layer, 7 neurons of hidden layer, and 1 neuron of output layer is the best for soil temperature time-series forecast, with a minimum sum squared error of 18.017. Relative errors between forecated value and measured value of the soil temperature all fall within 0～20%, with an average relative error of only 2.94%. The high range of relative error underscores the significance of BP neural network in daily soil temperature forecasting.