Abstract: Gravel-sand mulched field in northwestern arid area has a history of more than 300 years, but little is known about spatial-temporal dynamics of soil moisture, which is important for prediction of soil moisture. Therefore, in this study, we designed a field experiment to investigate dynamics of soil moisture in Experiment Station of Lanzhou University of Technology, Lanzhou, China. The experiment started from May 18 to August 3 in 2013. A field in width and length of 32 m was meshed by 4-m for soil sampling in each mesh grid. A total of 64 sampling points was designed. Soil samples were collected from the points at 0-10, 10-20, 20-30, and 30-50 cm for 12 times during the experiment. Based on the data, models were established for soil moisture prediction. Then, a similar experiment was conducted to collect samples to validate the model established in 2013. The results showed that the variability of surface soil moisture was largest and soil moisture distribution was in small patch-like pattern. The variability of soil moisture decreased with the increase of soil depth. Variation trend of mean soil moisture and coefficient of variation kept consistent over time in different soil layers. Mean soil moisture was 11.70% in 0-10 cm, 14.99% in 10-20 cm, 13.26% in 20-30 cm and 12.49% in 30-50 cm, respectively, indicating a upward trend with soil depth. However, the coefficient of variation was 0.2068 in 0-10 cm, 0.1821 in 10-20 cm, 0.1483 in 20-30 cm and 0.1241 in 30-50 cm, respectively, showing a downward trend with soil depth. The variation of soil moisture was largest in 0-10 cm. In addition, the study found that fragmentation degree of soil moisture distribution under dry condition was higher than wet condition, and temporal stability was high with the increase of soil depth. The relative deviation of soil moisture lied between -26.94% and 21.89% in 0-10 cm, between -23.40% and 17.26% in 10-20 cm, between -18.24% and 13.33% in 20-30 cm, between -15.73% and 11.60% in 30-50 cm. The standard deviation of the average relative deviation lied between 11.71% and 23.15% in 0-10 cm, between 10.85% and 22.18% in 10-20 cm, between 8.77% and 21.27% in 20-30 cm, between 6.96% and 17.35% in 30-50 cm. Spatial-temporal stability of soil moisture increased with the increase of soil depth. Based on information above, the representative measuring points of point 1, point 1, point 44, point 36 were selected for soil moisture estimation for 0-10, 10-20, 20-30 and 30-50 cm. The correlation coefficient between the estimated by representative measuring points and measured values was 0.877-0.968, and standard error and average deviation were small. Linear regression models were established based on the representative measuring points and validated by data obtained from 2014. The relative error between measured value and predicted value lied between 5.18% and 0.59% in 0-10 cm, between 15.42% and 0.55% in 10-20cm, between 1.07% and 10.33% in 20-30 cm, between -12.87% and 0.07% in 30-50 cm. It suggested that the prediction accuracy of models was high and the representative measuring points could be used for soil moisture estimation in the study area. The results provide valuble information for measuring point layout for soil moisture and accurate prediction of soil moisture in the study area.