Abstract: Predicting the location and water abundance of aquifers by drilling or geophysical methods before tunneling is very important to prevent water disasters in advance for the safety of coal mine production. The use of borehole transient electromagnetic method (BTEM) for advanced detection has obvious advantages. At present, the interpretation method is based on the calculated resistivity to qualitatively analyze rock water abundance, and it is impossible to predict the aquifer water abundance grade. So, the fully convolutional neural network (FCNN) method is proposed to predict the aquifer water abundance grade for BTEM. Firstly, according to the Archie formula, Kozeny-Carman formula, conductivity formula, and unit water inflow formula, the corresponding relationship between the resistivity per unit thickness of the sandstone and the four grades of aquifer water abundance grades according to the amount of water inflow per borehole is established. Then, the geological-geophysical model of different rock water abundance under the whole space condition is simulated by the three-dimensional finite difference time domain (FDTD) method, and the characteristic parameters used to predict the water abundance grades are extracted. To be close to the actual situation, 5%-15% random noise is added to the forward modeling results, and FCNN is used to train and simulate the prediction of aquifer water abundance. The average accuracy of aquifer water abundance grade predicted by the test set is 91.8%. Finally, the proposed method is used to predict the water abundance grade of coal seam after hydraulic fracturing in a mine. The results show that FCNN method is feasible and effective to predict the aquifer water abundance grade in the radial direction of the borehole and improve the detection accuracy of BTEM. This method will play an important role in the advanced detection of aquifer water abundance.