Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigating Map-Based Path Loss Models: A Study of Feature Representations in Convolutional Neural Networks

Ryan G. Dempsey, Jonathan Ethier, Halim Yanikomeroglu

TL;DR

This work investigates how CNN input representations affect map-based path loss prediction, specifically whether frequency and distance should be embedded as image channels or fed as scalar features. Using UK and Canada datasets with high-resolution DSM maps, the study finds that channel-based representations generally improve generalization, with the FINE configuration performing well when training data are diverse. FLIP, which places scalars in the regression layers, tends to overfit and show higher test RMSE, though ensemble approaches can mitigate some issues. The results offer practical guidance for designing robust CNN-based path loss estimators across regions, achieving around 7.25 dB RMSE on large, continent-spanning data sets.

Abstract

Path loss prediction is a beneficial tool for efficient use of the radio frequency spectrum. Building on prior research on high-resolution map-based path loss models, this paper studies convolutional neural network input representations in more detail. We investigate different methods of representing scalar features in convolutional neural networks. Specifically, we compare using frequency and distance as input channels to convolutional layers or as scalar inputs to regression layers. We assess model performance using three different feature configurations and find that representing scalar features as image channels results in the strongest generalization.

Investigating Map-Based Path Loss Models: A Study of Feature Representations in Convolutional Neural Networks

TL;DR

This work investigates how CNN input representations affect map-based path loss prediction, specifically whether frequency and distance should be embedded as image channels or fed as scalar features. Using UK and Canada datasets with high-resolution DSM maps, the study finds that channel-based representations generally improve generalization, with the FINE configuration performing well when training data are diverse. FLIP, which places scalars in the regression layers, tends to overfit and show higher test RMSE, though ensemble approaches can mitigate some issues. The results offer practical guidance for designing robust CNN-based path loss estimators across regions, achieving around 7.25 dB RMSE on large, continent-spanning data sets.

Abstract

Path loss prediction is a beneficial tool for efficient use of the radio frequency spectrum. Building on prior research on high-resolution map-based path loss models, this paper studies convolutional neural network input representations in more detail. We investigate different methods of representing scalar features in convolutional neural networks. Specifically, we compare using frequency and distance as input channels to convolutional layers or as scalar inputs to regression layers. We assess model performance using three different feature configurations and find that representing scalar features as image channels results in the strongest generalization.
Paper Structure (10 sections, 2 figures, 4 tables)

This paper contains 10 sections, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Training and Evaluation Data
  3. United Kingdom Data
  4. Additional Test Data
  5. Channel Configurations
  6. Model Performance
  7. Cross-Validation in the United Kingdom
  8. No-Holdout Models and Blind Testing in Canada
  9. Investigating FLIP Performance
  10. Concluding Remarks

Figures (2)

  • Figure 1: Convolutional neural network topology. *Original and FINE both use a filled frequency channel, with FINE replacing the original 2-D distance grid with a filled channel containing 3-D link distance. FLIP represents both frequency and distance as scalar inputs to the FCN.
  • Figure 2: Average loss curve across all 10 no-holdout models.