The digital elevation model (DEM) serves as a vital data source for surface 3D modeling. Due to the limitations in sampling conditions and cost constraints, we usually obtain unevenly distributed and relatively sparse sampling points. To reconstruct a complete DEM of the sampling area, we need to utilize spatial interpolation algorithms. However, traditional spatial interpolation methods typically have lower model complexity and often involve a large number of iterative calculations to approximate the points to be interpolated. This often results in significant interpolation errors and low real-time performance. We propose a multi-scale conditional generative adversarial network (multi-scale cGAN) with adaptive joint loss weights. In addition, during the model training process, we design a joint loss function that incorporates generator adversarial loss, content loss, and perceptual loss, with the ability to adaptively adjust the weight coefficients of each component, thereby optimizing model training and further improving its generalization and generation ability. The experimental results demonstrate that compared with the traditional spatial interpolation algorithm and other typical deep learning–based models, the interpolation error on typical land DEM data (including slopes, valleys, and ridges) is smaller, and the interpolated image has the highest clarity and similarity compared with the original image. Overall, our approach demonstrates high robustness and low error when dealing with DEM spatial interpolation tasks in complex terrain environments while also possessing the potential for expansion into various terrain environment DEM reconstruction applications.