Accurate knowledge of the three-dimensional ionospheric electron-density distribution is essential for reliable radio-wave propagation modeling, yet global empirical models (e.g., IRI-2020, NeQuick2) fail to capture local and short-term variability. In this work, we propose a U-Net–based super-resolution CNN (SRCNN) that reconstructs a regionally specialized 3D electron-density distribution over the Korean Peninsula from sparse, high-fidelity input profiles. These input profiles are generated for two sites (Icheon and Jeju) by combining direct ionosonde measurements for the bottomside with an ionosonde-corrected IRI-2020 model for the topside. The AI model was trained on electron-density distributions produced by the IRI-2020 model. The proposed model demonstrates significant improvements over the standard IRI model, showcasing its stability across all solar activity levels. Most notably, under solar-maximum conditions, the root mean square relative error (RMSRE) was drastically reduced at Icheon (from 367.23% to 16.04%) and Jeju (from 538.12% to 9.68%). The model also consistently improved other key metrics, such as the F2-peak altitude error and the Pearson correlation coefficient (r > 0.99), proving its robust performance. The proposed approach can contribute to improving ionospheric error correction and signal quality in precise GNSS positioning, space surveillance radar, and satellite communication systems.