---
license: cc-by-4.0
task_categories:
- image-segmentation
- depth-estimation
tags:
- remote-sensing
- satellite
- clouds
- lidar
- icesat-2
- meteosat
- era5
- meteorology
- semi-supervised-learning
pretty_name: CloudTrack-3D
size_categories:
- 1K
---
## Visual Examples from the Benchmark
### 1. Co-Registered Satellite Imagery & 3-Beam ICESat-2 Lidar Curtain
Each granule pairs full-disk sector **IR 10.5 μm** brightness temperature and **VIS 0.6 μm** solar reflectance with 3 parallel ICESat-2 lidar ground tracks (`gt1l`, `gt2l`, `gt3l`) providing vertical cloud top and base boundaries up to 10 atmospheric layers:
### 2. Dense 2D Cloud Top & Base Reconstruction vs. Along-Track Lidar Ground Truth
Dense 2D predictions of cloud mask, cloud top height (`0–14 km`), and cloud base height (`0–14 km`) compared against the sparse ICESat-2 lidar curtain:
---
## Dataset Splits & Protocols
| Split Protocol | Total Granules | Train Granules | Val/Test Granules | Eval Patches (256 × 256) | Modalities Included |
| :--- | :---: | :---: | :---: | :---: | :--- |
| **`allday_490` (1f IR All-Day)** | 490 | 392 | 98 | 588 | IR 10.5 μm (t = 0) + Lat/Lon + ERA5 + ICESat-2 3-Beam |
| **`multispectral_3frames_408` (3f IR+VIS)** | 408 | 327 | 81 | 486 | IR 10.5 μm (t = 0, -15m, -30m) + VIS 0.6 μm (t = 0, -15m, -30m) + Lat/Lon + ERA5 + ICESat-2 |
| **`synoptic_block_408` (14-Day Block + 48h Buffer)** | 408 | 304 | 91 *(+13 buffer)* | 546 | Weather-regime-disjoint evaluation split (±48 h exclusion buffer between blocks) |
---
## Repository Files & Archives
| File / Directory | Size | Description |
| :--- | :---: | :--- |
| `cloudtrack3d_splits.json` | `3.1 MB` | Complete split manifest (`allday_490`, `multispectral_3frames_408`, `synoptic_block_408`) with exact `256 × 256` evaluation patch bounding boxes |
| `granule_manifest_train.csv` / `granule_manifest_test.csv` | `113 KB` | Flat CSV metadata tables of all 490 matched granules (timestamps, 3-frame filenames, ERA5 links, split flags, footprint counts) |
| `viewer/train/` & `viewer/test/` | `~18 MB` | Side-by-side (**IR 10.5 μm \| VIS 0.6 μm**) preview images rendered directly in the Hugging Face Dataset Viewer above |
| `vfms_490.tar` | `205 MB` | All 490 matched ICESat-2 `ATL09` L3A `.h5` lidar granules (`vfms/ATL09_*.h5`) |
| `era5_matched_490.npz` | `55 MB` | Pre-extracted ERA5 atmospheric profiles (`18 × 61 × 77` `float16`: Temperature and Relative Humidity across 9 pressure levels `200–900 hPa`) keyed by `granule_id` |
| `ir105_matched_1467.tar` | `1.65 GB` | 1,467 matched 3-frame (`t = 0, -15m, -30m`) full-resolution `4096 × 4096` **IR 10.5 μm** images (`ir10.5/YYYY-MM-DD_HH-MM.png`) |
| `vis06_matched_1252.tar` | `4.13 GB` | 1,252 matched 3-frame (`t = 0, -15m, -30m`) full-resolution `4096 × 4096` **VIS 0.6 μm** images (`vis0.6/YYYY-MM-DD_HH-MM.png`) |
| `METEOROLOGY.md` | `12 KB` | Physical documentation of the satellite channels, lidar profiles, and ERA5 pressure levels |
---
## Quick Start: Downloading & Unpacking
```bash
# 1. Download the dataset repository
huggingface-cli download cloudTrack3D/CloudTrack-3D --repo-type dataset --local-dir ./data
cd ./data
# 2. Unpack the full-resolution (4096 x 4096) satellite and lidar archives
tar -xf vfms_490.tar
tar -xf ir105_matched_1467.tar
tar -xf vis06_matched_1252.tar
```
### Loading ERA5 Profiles & Split Manifest in Python
```python
import json
import numpy as np
# Load train/test splits and exact 256x256 evaluation patch coordinates
with open("data/cloudtrack3d_splits.json") as f:
splits = json.load(f)
train_granules = splits["splits"]["multispectral_3frames_408"]["train"]
test_granules = splits["splits"]["multispectral_3frames_408"]["test"]
# Load pre-extracted 18-channel ERA5 profiles (9 Temperature + 9 Relative Humidity levels, 200-900 hPa)
era5_npz = np.load("data/era5_matched_490.npz")
sample_profile = era5_npz[train_granules[0]["granule_id"]] # shape: (18, 61, 77)
```