Abstract: Physics-Informed Neural Networks (PINN) models have been widely used for forward solving of groundwater level and quantity problems. However, for hydrogeological parameter inversion problems, the use of PINN model alone often has a large uncertainty due to the fact that different parameter distributions could produce the same result. Although researchers have proposed many methods such as slug test and data assimilation, the small sample problem has always plagued the global parameter field description in practice. This research focused on the advanced generative deep learning method, in order to solve the problem of hydrogeological parameter inversion with small samples. In this study, PINN and Conditional Generative Adversarial Networks (CGAN) are coupled to form PICGAN (Physics-Informed Conditional Generative Adversarial Networks, PICGAN) model. And a two-dimensional non-homogeneous non-stationary arithmetic model has been created to test the effectiveness of PICGAN. A heterogeneous hydraulic conductivity field was set up as the inversion target. The numerical solution of groundwater level was obtained as the reference true value. Under the jointed constraints of physical conditions and reference true groundwater level, the discriminator of the PICGAN model continuously required the generator to update the global hydrogeological parameter field more in line with the reality until the convergence criteria of the generator and the discriminator were satisfied. At this point, it could be considered that the PICGAN model has completed the inverse solution of the hydraulic conductivity field of the non-stationary non-homogeneous confined aquifer, and could also simultaneously simulate and predict the water level. The results showed that in the case simulation with 5% sampling rate, the root-mean-square error of water head simulated by the PICGAN model could be stabilized at about 0.95 m, with an accuracy of 89%. The root-mean-square error of hydraulic conductivity field could be stabilized at about 0.69 m/d, with an accuracy of up to 95%, and the distribution form was highly consistent with the reference field. As the sampling rate increased, the global error of the model would be further reduced. After the sampling rate reaches 10% or more, the inversion error of the global hydraulic conductivity was reduced to 0.35 m/d, with an accuracy of 97%. In conclusion, the PICGAN model proposed in this study can provide new ideas and methods for bidirectional solution of groundwater problems under small sample conditions, especially for the inversion of heterogeneous hydrogeological parameter fields.