In the process of deep foundation pit construction, a small amount of borehole data is needed to reconstruct the 3D model of the soil layer to obtain the distribution of soil information. This paper proposes a method of soil layer reconstruction based on machine learning. First, a soil layer generation algorithm is designed to enhance the data of the soil layer training data set. Then, the prediction model feature coding method is designed according to the data structure of the borehole information. As the standard input of the prediction model, the convolutional neural network model is built to extract the features of the soil layer structure to form the soil layer prediction model. Subsequently, the prediction model is used to predict the soil layer attributes of the discrete grid points in the predicted block to obtain soil layer data. Finally, the Marching Cubes algorithm is used to generate closed isosurfaces for the soil layer data to form solid blocks of the soil layer, thereby realizing the reconstruction of the three-dimensional soil layer. This model can adapt to different layers and different types of strata, and has the preliminary conditions for practical engineering applications.