Replacing Data with Models for REM Estimation

This repository implements a modular, two-stage deep learning pipeline for estimating Radio Environment Maps (REM) from monocular aerial imagery.

By replacing physical LiDAR-derived elevation data with a learned estimator, this approach eliminates the need for expensive sensing infrastructure at inference time, offering a scalable and memory-efficient alternative for urban radio propagation modeling.

Project Architecture

The pipeline is split into two distinct stages.

Stage 1: Elevation Estimation (Im2Ele)

Based on the SENet architecture, this stage performs monocular height reconstruction.

1. Input: Single RGBI aerial image.
2. Output: Predicted normalized Digital Surface Model (nDSM).
3. Location: See the IMELE-master directory for the retraining notebook.

Stage 2: Radio Map Learning (RML)

Utilizes a lightweight LitRadioUNet to predict signal propagation.

1. Input: Aerial image, predicted nDSM (from Stage 1), transmitter location, and antenna gain.
2. Output: Spatial path-loss map.
3. Location: See the RML directory for the training and evaluation notebooks.

Installation & Setup

The requirements for the Im2Ele model is in the notebook from which we retrain or with the "IMELE-master/requirements.txt".

For Radio Map Model we recommend using the conda environment.

cd conda_env
conda env create -f env_complete.yml --solver=classic
source activate rmbd_env
python -m ipykernel install --user --name rmbd_env --display-name "Python (rmbd_env)"

Workflow

• Retrain Stage 1: Use the notebook in IMELE-master to train the height estimator on the RMDirectionalBerlin dataset.
• Generate Height Maps: Use the provided generation notebook to produce predicted nDSMs for the entire dataset.
• Train Stage 2: Use the notebook in RML to train the Radio Map model using the predicted heights from Stage 1 as input features.

Dataset & Credits

This research utilizes the RMDirectionalBerlin dataset. We acknowledge the authors of the original architectures used in this pipeline:

• RML (Radio Map Learning)
• Im2Ele (Image to Elevation)