Interferometric Synthetic Aperture Radar (InSAR) stands as a rapidly advancing space-borne Earth observation technology, which finds extensive application in monitoring domains such as surface subsidence in mining areas. Nevertheless, most metal mine areas are located in densely vegetated regions, and factors such as vegetation cover cause InSAR decoherence, which has become a major technical bottleneck. The occurrence of this issue leads to a reduction in interferometric points, thereby impairing the accuracy of monitoring. Based on the settlement value of known interference points, this paper fused convolutional neural network ( CNN ) and radial basis function ( RBF ) for spatial interpolation, adopted 80% training and 20% verification of random split data to reasonably predict the settlement of non-interference point area, and compared it with two traditional interpolation methods: spline interpolation and Kriging interpolation. The optimal interpolation method was obtained through error analysis. The results show that the effect of spline interpolation and Kriging interpolation is not ideal ; CNN interpolation can accurately predict the surface subsidence without interference points due to its ' data-driven ' characteristics. After incorporating RBF fitting, the errors generated by each interpolation method decrease significantly, which attests to a marked improvement in interpolation performance; Both before and after the integration of the RBF, the interpolation performance of the CNN method is notably superior to that of the other two methods; The application of appropriate interpolation methods can alleviate the problem of sparse interferometric points efficiently, yielding more accurate and comprehensive surface subsidence data. The study overcomes the limitations of single interpolation methods in terms of accuracy for subsidence prediction in sparse interferometric points areas, providing a more efficient technical approach for subsidence value completion.