Red neuronal híbrida U-net-INR para la reconstrucción de mapas electromagnéticos dispersos de alta precisión

Sparse sampling hinders high-precision electromagnetic environment map reconstruction, crucial for spectrum management and situation awareness. To overcome the limitations of traditional interpolation and single-model deep learning, we propose a hybrid U-Net-Implicit Neural Representation (INR) algorithm. This end-to-end framework integrates U-Net for semantic feature extraction and INR for continuous field mapping, effectively restoring nonlinear propagation details and ensuring physical consistency. Tests at sparsity levels of 0.005-0.05 demonstrate that our method achieves superior accuracy (lower MSE/MAE, higher PSNR) compared to Kriging and inverse distance weighting (IDW) interpolation. By enabling high-fidelity reconstruction from extremely sparse data, this work provides a robust solution for applications requiring accurate electromagnetic mapping under practical sampling constraints.