A Deep Split-Step Wavelet Model for the Long-Range Propagation

This article presents a new approach based on a deep-learning method using a U-Net architecture to generate electromagnetic propagation over a specific terrain. For this purpose, the learning dataset is constructed artificially using a fast split-step wavelet (SSW) method. For this phase, the synthetic 1D profiles are randomly generated from rectangle and triangle shapes. This latter allows for conveying the “staircase” model used in SSW. To ensure a precise sampling of the underlying manifold, the study employs Latin Hypercube Sampling. To achieve robust and precise predictions, a specific loss function is proposed. To evaluate this approach, numerical tests are realized. These tests demonstrate the effectiveness of the proposed method in realistic terrain.