DRaGon: Mining Latent Radio Channel Information from Geographical Data Leveraging Deep Learning

TL;DR

DRaGon introduces a machine learning approach that leverages geographical data and deep learning to generate Radio Environmental Maps, significantly improving path loss prediction accuracy over traditional models.

Contribution

It presents a novel deep learning-based method for automatic radio channel modeling using geographical data, enhancing accuracy beyond existing analytical and ray tracing techniques.

Findings

DRaGon outperforms traditional models in path loss prediction accuracy.

Combining domain expertise with deep learning improves modeling results.

The approach is validated against real measurements and established methods.

Abstract

Radio channel modeling is one of the most fundamental aspects in the process of designing, optimizing, and simulating wireless communication networks. In this field, long-established approaches such as analytical channel models and ray tracing techniques represent the de-facto standard methodologies. However, as demonstrated by recent results, there remains an untapped potential to innovate this research field by enriching model-based approaches with machine learning techniques. In this paper, we present Deep RAdio channel modeling from GeOinformatioN (DRaGon) as a novel machine learning-enabled method for automatic generation of Radio Environmental Maps (REMs) from geographical data. For achieving accurate path loss prediction results, DRaGon combines determining features extracted from a three-dimensional model of the radio propagation environment with raw images of the receiver area within a deep learning model. In a comprehensive performance evaluation and validation campaign, we compare the accuracy of the proposed approach with real world measurements, ray tracing analyses, and well-known channel models. It is found that the combination of expert knowledge from the communications domain and the data analysis capabilities of deep learning allows to achieve a significantly higher prediction accuracy than the reference methods.