Kriging interpolation is an important geostatistical method, but it is constrained by data. In areas with high data density, the interpolation results are reliable, but the farther the distance from the control point, the greater the error of the interpolation results. In order to further improve the accuracy of interpolation calculation, the mixed density probability distribution data prediction method can be used. First, a low-frequency model is obtained by interpolating the conventional kriging interpolation method, and then the mixed density probability network is trained on the known information such as logging geology and other data, and then the trained network is used to predict other unknown location data, to get a high frequency model. Finally, the two results are fused with high and low frequencies to obtain the final geological model. A near-surface modeling test is carried out on a certain work area, and it is proved that the method can effectively improve the interpolation accuracy.