Diffraction and Scattering Aware Radio Map and Environment Reconstruction using Geometry Model-Assisted Deep Learning

TL;DR

This work tackles 3D radio map construction without city maps by introducing a geometry-aware neural network that jointly learns a virtual obstacle map and a 6D radio map from RSS data. It decomposes propagation into blockage, diffraction, and local scattering, with a LOS/D0-based masking, a transformer-based diffraction module, and a rotation/scale-invariant scattering branch powered by a spatial transformer. The approach reconstructs plausible virtual environments, improves radio-map accuracy by 10–18% over baselines, and demonstrates strong transferability to new environments and altitudes, plus a UAV relay placement benefit with dramatically reduced search effort. Overall, the method advances practical radio planning and environment-aware wireless modeling in scenarios where city maps are unavailable or outdated.

Abstract

Machine learning (ML) facilitates rapid channel modeling for 5G and beyond wireless communication systems. Many existing ML techniques utilize a city map to construct the radio map; however, an updated city map may not always be available. This paper proposes to employ the received signal strength (RSS) data to jointly construct the radio map and the virtual environment by exploiting the geometry structure of the environment. In contrast to many existing ML approaches that lack of an environment model, we develop a virtual obstacle model and characterize the geometry relation between the propagation paths and the virtual obstacles. A multi-screen knife-edge model is adopted to extract the key diffraction features, and these features are fed into a neural network (NN) for diffraction representation. To describe the scattering, as oppose to most existing methods that directly input an entire city map, our model focuses on the geometry structure from the local area surrounding the TX-RX pair and the spatial invariance of such local geometry structure is exploited. Numerical experiments demonstrate that, in addition to reconstructing a 3D virtual environment, the proposed model outperforms the state-of-the-art methods in radio map construction with 10%-18% accuracy improvements. It can also reduce 20% data and 50% training epochs when transferred to a new environment.

Figures (9)

• Figure 1: Modeling the diffraction between point $A_{0}$ and point $B$ by exploiting the geometry relation of distances $d_{i}$ and angles $\theta_{i}$ from the virtual obstacles (gray cubes).
• Figure 2: While the propagation environments from the TX to three RXs are different (left figure), the local geometry structures are similar after proper rotation and scaling (right figures). Thus, the same propagation mechanism should be learned.
• Figure 3: The overall architecture to exploit the geometry structure of the environment for joint 6D radio map and virtual environment map reconstruction.
• Figure 4: The transformer to learn the diffraction mechanism by exploiting the geometry structures of the diffraction distances $d_{i}$ and angles $\theta_{i}$.
• Figure 5: The CNN to learn the geometry structure of the local environment by exploiting the rotation invariance and scale invariance properties.