Map-Based Path Loss Prediction in Multiple Cities Using Convolutional Neural Networks

Radio deployments and spectrum planning benefit from path loss predictions. Obstructions along a communications link are often considered implicitly or through derived metrics such as representative clutter height or total obstruction depth. In this paper, we propose a path-specific path loss prediction method that uses convolutional neural networks to automatically perform feature extraction from 2-D obstruction height maps. Our methods result in low prediction error in a variety of environments without requiring derived metrics.

💡 Research Summary

The paper presents a data‑driven approach for predicting radio path loss that leverages high‑resolution geographic information system (GIS) data in the form of digital surface model (DSM) height maps. Unlike traditional models such as ITU‑R P.1812‑6, which rely on scalar descriptors like representative clutter height or total obstruction depth, the authors feed the raw 2‑D obstruction height profile directly into a convolutional neural network (CNN).

Two primary datasets are used: (1) drive‑test measurements from the UK Office of Communications (Ofcom) covering six frequency bands (449 MHz to 5850 MHz) and seven cities, and (2) the UK open digital elevation model (DEM) providing 1 m‑resolution DSM data for six of those cities. After filtering for signal‑to‑noise ratio (> 6 dB) and link distance (> 50 m), about 4.2 million measurements remain. For each link, a rectangular grid of width W = 61 m and length equal to the integer link distance (in meters) is extracted from the DSM, corrected for Earth curvature, and then resampled to a uniform length of 256 pixels. Four input channels are constructed: (i) normalized frequency (constant per sample), (ii) Euclidean distance from the transmitter to each pixel, (iii) the DSM height values, and (iv) a “direct‑path” channel that encodes the antenna heights along the centre line of the grid. Heights are normalized per‑sample to