High-resolution digital elevation model（DEM） in large districts is difficult to be acquired due to the limitation of cost and technology. Usually, it can be obtained by super-resolution reconstruction（downscale） from low-resolution DEM. However, the accuracy of the DEM generated by conventional downscale methods is insufficient. With the development of image downscale, convolutional neural network（CNN） has achieved success. To improve DEM accuracy, a very deep convolutional networks super-resolution method（VDSR）was designed to reconstruct the terrace DEM with obvious undulation characteristics. The deep neural network was used to learn nonlinear mapping between high-resolution DEM and low-resolution DEM, at the same time, residual learning method were used to reduce training difficulty. In order to compare, bicubic interpolation method, sparse mixed estimation method and VDSR method were used to reconstruct the DEM and slope. The slope data were extracted from the DEM results. The mean value of DEM difference of three methods were 0.41m, 0.34m and 0.34m, respectively. The RMSE of DEM were 0.5945m, 0.5715m and 0.4869m, respectively. The mean value of slope difference of three methods were 3.02°, 2.04° and 1.99°, respectively. The RMSE of slope were 3.6498°, 3.1360° and 2.7387°, respectively. The running time were 0.052s, 663.39s and 2.16s, respectively. By comprehensive comparison, for 10m, 20m and 40m DEM, the result showed that VDSR method had great advantage in spatial distribution, error and running time, and it was suitable for super-resolution reconstruction in areas with complex terrain such as terrace.