While physics-informed neural networks (PINNs) can solve the problem pertaining to the absence of boundary conditions in soil water systems, their results exhibit low accuracy primarily due to insufficient utilization of the available prior knowledge regarding soil hydraulic parameters. In this research, an improved PINNs framework is proposed, which introduces an optimizable saturated hydraulic conductivity into the activation function, and an advanced optimization strategy is developed to identify the optimal superparameters for the multi-objective loss function. The PINNs was trained using synthetic volumetric soil water content (VSWC) and soil matric potential (SMP) data generated by a numerical solution of the Richardson-Richards equation (RRE) for three soil types (silt loam, loam and sandy loam). The results show that the proposed framework increases the accuracy of SMP estimations in the unsaturated soil system. The results reveal that the relative error achieved by the proposed framework in loam or silt loam has been reduced by two orders of magnitude in comparison with that achieved by the framework introduced by Bandai and Ghezzehei (2020), indicating a significant improvement. While there is a slight reduction in the accuracy of volumetric soil water content estimation, this minor reduction has minimal practical significance. Both the soil water retention curve and the soil hydraulic conductivity exhibit superior performance at the near-saturation scale. For unsaturated flow in homogeneous soil, the proposed PINNs framework provides accurate estimations of soil hydraulic parameters and holds significant potential for the practical application and widespread adoption of PINNs in the realm of soil hydrodynamics.