jliang097/FARM_training_test

FARM Aerial Radio Map (ARM) Dataset

Paper:

• FARM: Foundational Aerial Radio Map for Intelligent Low-Altitude Networking (https://arxiv.org/abs/2604.17362)

Overview

This repository releases the constructed ARM datasets based on ARM-Omni for FARM training, in-domain evaluation (D1-D10), and zero-shot evaluation (P1, F1, and A1). The dataset coverage is summarized below:

Dataset	Frequencies (GHz)	Max Rx Height (m)	Beamwidths	Map Grid Size	Volume
D1	2.1, 3.3, 5.9	120	30°, 120°, Iso	512 × 512	15000
D2	3.5, 4.9, 5.9	130	30°, 120°, Iso	512 × 512	15000
D3	2.1, 4.9, 5.9	110	30°, 120°	512 × 512	15000
D4	3.3, 3.5, 4.9	110	30°, 120°, Iso	512 × 512	15000
D5	3.3, 4.9, 5.9	120	30°, 120°, Iso	512 × 512	15000
D6	2.1, 3.3, 3.5	130	30°, 120°	512 × 512	15000
D7	2.1, 3.3, 3.5, 5.9	130	30°, Iso	512 × 512	15000
D8	2.1, 3.5, 4.9, 5.9	100	120°, Iso	512 × 512	15000
D9	2.1, 3.3, 3.5, 4.9	120	30°, 120°	512 × 512	15000
D10	2.1, 3.3, 3.5, 4.9, 5.9	130	30°, 120°, Iso	512 × 512	15000
P1	2.1, 3.3, 3.5, 4.9, 5.9	150	30°, 120°, Iso	256 × 256	15000
F1	2.6, 7.1	120	30°, 120°, Iso	512 × 512	15000
A1	2.1, 3.3, 3.5, 4.9, 5.9	120	60°	512 × 512	15000

A fuller ARM-Omni release is planned for a future update.

Repository Layout

Each scene is stored as a compressed archive on the Hub:

<dataset>/scene_<scene_id>.tar.gz

After extraction, the scene contains .npy ARM tensors organized as:

<dataset>/<scene_id>/<frequency>/<antenna_pattern>/tx<ID>_yaw<VALUE>.npy

Example:

D1/7/freq59/iso/tx45_yaw0.npy

Scene IDs are not globally continuous across dataset groups. Users should rely on the extracted folder names as the authoritative scene IDs.

Data Format

Each released raw .npy file stores an ARM volume with shape:

1 × 30 × 512 × 512

The dimensions correspond to one radio-map channel, 30 height layers, and a 512 × 512 spatial grid. During data loading, different input sizes can be obtained by selecting a subset of height layers and applying a spatial crop centered on the transmitter pixel. For example, this procedure is used to obtain the 256 × 256 crop for P1.

Signal Encoding and Scene Metadata

• The no-signal truncation threshold is -120 dB. In the encoded uint8 representation, this threshold corresponds to pixel value 130.
• Building occupancy is embedded in the released ARM tensors; in each height-layer radio map, pixels with value 0 correspond to building-occupied regions.
• The transmitter location metadata is stored in the scene-level CSV file:

<dataset>/<scene_id>/tx_positions_512.csv

## Citation

If you use this dataset, please cite our paper:

```text
S. Gao, J. Liang, Y. Yuan, W. Lu, G. Shen, and L. Yang, "FARM: Foundational Aerial Radio Map for Intelligent Low-Altitude Networking," arXiv preprint arXiv:2604.17362, 2026.

@article{gao2026farm,
  author  = {Gao, Shijian and Liang, Jiahui and Yuan, Yifeng and Lu, Wenlihan and Shen, Guobin and Yang, Liuqing},
  title   = {{FARM}: Foundational Aerial Radio Map for Intelligent Low-Altitude Networking},
  journal = {arXiv preprint arXiv:2604.17362},
  year    = {2026}
}