Abstract: Abstract: It is very important to improve water use efficiency and achieve precision irrigation that soil moisture content is predicted accurately. The spatial and temporal variability of soil moisture content is complex because soil moisture content can be affected by various factors, such as soil properties, plant, and environment. The time-series field soil moisture data were nonlinear. There was strong relationship of soil moisture content between the adjacent depths. The work presented in this paper aimed to contribute to predicting soil moisture content in different depths by proposing a time-delay neural network (TDNN). TDNN is an artificial neural network model in which all the neuron-like units (nodes) are fully connected by directed connections. Each unit has a time-varying real-valued activation and each connection has a modifiable real-valued weight. It has 3 layers: input layer, hidden layer and output layer. In this paper, a prediction model based on the TDNN was presented to predict soil moisture content in 6 field depths (10, 20, 30, 40, 50 and 70 cm). The framework of the prediction model based on the TDNN included input layer with 6 units, hidden layer with 10 units and output layer with 6 units. Three training algorithms, which were Levenberg-Marquardt (L-M) method, conjugate gradient method and momentum increase method, were tested. The simulation results show that the L-M method was the best, followed by the conjugate gradient method, and the momentum increase method was the worst. The original experimental data were acquired from maize field in Shangzhuang experiment station, China Agricultural University in Beijing. The sample data set of soil moisture content prediction model based on the TDNN was generated from the original data set, which was calibrated by drying method and preprocessing algorithm. The prediction accuracy of the soil moisture content prediction model was influenced greatly by the training sample's size. The experiment results showed that the best training accuracy could be gotten when the number of training samples was more than 40% of that of all samples. The prediction model would over fit when the number of training samples was more than 50%. In this paper, every 20 sample data were divided into a set and 9 data were set as training data and other 11 data as test data. The sample data set of soil moisture content prediction model was divided into training set and test set. The results showed that the relative error of soil moisture content prediction for 10 and 20 cm was less than 7%. As for 30, 40, 50 and 70 cm, the relative error was less than 4.5%. The prediction model based on the TDNN presented in this paper can be used to predict the soil moisture content in different depth and can be treated as a solution to grasp the distribution of soil moisture content