The Moho is the interface between crust and mantle, and accurate location of the Moho is important for both resource exploration and deep earth condition and structural change investigations. The Parker-Oldenburg (P-O) method, is simple and efficient and thus has been extensively applied in the frequency domain and for Moho depth inversion. However, Moho fluctuation simulations using the P-O method are not reliable because of the lack of field geographic data constraints during the inversion process and excessively smoothing of data details caused by using a filter to correct the source data signals. To solve those problems, we propose an improved iteration P-O method with variable density, the iterative process is constrained by geological data in the inversion parameters, and the variable depth of the gravity interface is iterated using an equivalent form of upward continuation in the Fourier domain, which is more stable and convergent than downward continuation term in original P-O method. Synthetic experiments indicate that improved method has the better consistency among the simulations than original method, and our improved method has the smallest RMS of 0.59 km. In a real case, we employed the improved method to invert the Moho depth of the South China Sea, and the RMS between our Moho model and the seismological data is the smallest value of 3.87 km. The synthetic experiments and application of the model to the SCS further prove that our method is practical and efficient.