ImpactMesh-X

ImpactMesh-X extends the original ImpactMesh (Flood / Fire) dataset along two axes: time and robustness. It benchmarks hazard mapping models on long-term multi-temporal context and atmospheric occlusion, rather than single-image, cloud-free snapshots.

The dataset covers two hazards, Flood and Fire, with 19,891 patches total (256×256 px, 10 m resolution). Every patch carries three modalities:

• S1RTC -- Sentinel-1 RTC, 2 bands (VV, VH), 16 timesteps
• S2L2A -- Sentinel-2 L2A, 12 bands (B01–B09, B8A, B11, B12), 16 timesteps, in three cloud-cover variants (clear / cloudy / realistic)
• DEM -- Copernicus WorldDEM-30, 1 band (elevation), static

...plus a hazard annotation mask and per-patch metadata (event date, CEMS event id, AOI, geolocation, pool). Timesteps span a ±180-day window around the event date. Positive (hazard-present) and negative (hazard-free control) patches are mixed in the same folders and distinguished by the pool column.

Hazard	Patches (total)	Positive	Negative	CEMS events	Event dates
Flood	9,723	4,410	5,313	191	2016 – 2026
Fire	10,168	4,801	5,367	215	2016 – 2026

For dataset loading code, baselines, and the full paper training pipeline, see the GitHub repository.

---

Examples

Each patch's Sentinel-1 and three Sentinel-2 cloud variants (clear / realistic / cloudy), shown around the event window, alongside the DEM and ground-truth annotation.

Fire — EMSR214_1_32TNN_x534005_y4719165 Flood — EMSR838_1_43SDS_x428395_y3603665

---

Data Structure

ImpactMesh-X/
├── Flood/
│   ├── S1RTC/
│   │   └── {patch_id}_S1RTC.zarr.zip
│   ├── S2L2A/
│   │   ├── clear/{patch_id}_S2L2A.zarr.zip
│   │   ├── cloudy/{patch_id}_S2L2A.zarr.zip
│   │   └── realistic/{patch_id}_S2L2A.zarr.zip
│   ├── DEM/
│   │   └── {patch_id}_DEM.tif
│   ├── MASK/
│   │   └── {patch_id}_annotation_flood.tif
│   └── flood_patch_meta.parquet
├── Fire/
│   ├── S1RTC/{patch_id}_S1RTC.zarr.zip
│   ├── S2L2A/{clear,cloudy,realistic}/{patch_id}_S2L2A.zarr.zip
│   ├── DEM/{patch_id}_DEM.tif
│   ├── MASK/{patch_id}_annotation_wildfire.tif
│   └── fire_patch_meta.parquet
└── splits/
    ├── flood_splits.json     # official train/val/test patch_ids (positive samples only)
    ├── flood_norm.json       # per-band mean/std, computed on the train split
    ├── fire_splits.json
    └── fire_norm.json

Each modality folder is flat, one file per patch_id, joinable against {hazard}_patch_meta.parquet. splits/ mirrors the official split from the GitHub repository.

S1RTC/S2L2A are zipped Zarr stores with the full 16-timestep stack:

<zarr.Group> {patch_id}_S2L2A.zarr.zip
Dimensions:  (time: 16, band: 12, y: 256, x: 256)
Arrays:
    bands        (time, band, y, x) int16    # band order given by `band`
    band         (band,)           <U3       # ["B01","B02",...,"B8A","B09","B11","B12"]
    time         (time,)           int64
    cloud_mask   (time, y, x)      uint8     # S2L2A only
    SCL          (time, y, x)      uint8     # S2L2A only
    nan_mask     (time, y, x)      int8
Attributes: patch_id, event_date, collection, resolution, epsg, ...

S1RTC is the same, with band = ["vv", "vh"] and no cloud_mask/SCL. DEM and MASK are single-band GeoTIFFs. MASK is binary for Fire (0/1=wildfire) and 3-class for Flood (0/1=permanent water/2=flooded, the positive class); both may contain -1 for ignored pixels.

---

Download & Quickstart

# 1. Clone the code repo
git clone https://github.com/PaulH97/ImpactMesh-X.git
cd ImpactMesh-X

# 2. Install the HF CLI
pip install --upgrade huggingface_hub

# 3. Authenticate (only first time)
hf auth login  # paste your token from https://huggingface.co/settings/tokens

# 4. Pull one hazard plus the official split, e.g. Flood
mkdir -p data
hf download paulhoehn/ImpactMesh-X --repo-type dataset --local-dir data --include "Flood/*" "splits/flood_*"

# ...or download everything
hf download paulhoehn/ImpactMesh-X --repo-type dataset --local-dir data

Each modality folder ships as one or more .tar archives -- unpack them before running the code:

# from repo root, for each modality archive under data/<hazard>/:
for archive in data/Flood/*/*.tar data/Flood/*/*/*.tar; do
  [ -f "$archive" ] && tar -xf "$archive" -C "$(dirname "$archive")" && rm "$archive"
done

Then install the loader's dependencies and run the example:

pip install -r requirements.txt
python example.py --root data/Flood --task flood

from dataset import ImpactMeshDataset

ds = ImpactMeshDataset(root="data/Flood", task="flood", pool="positive")
sample = ds[0]  # {"image": {...}, "mask": ..., "metadata": {...}}

See the GitHub repository for a runnable example, workflow, and reproducing the paper's training pipeline.

---

Acknowledgement

Sentinel-2 Level-2A data were downloaded from Microsoft Planetary Computer and are provided under Copernicus Sentinel license conditions (© European Union 2015–2025, ESA) (https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a).

Sentinel-1 Radiometrically Terrain Corrected (RTC) SAR data were retrieved from Microsoft Planetary Computer (calibrated to GRD and terrain-corrected using PlanetDEM) under Copernicus Sentinel license terms (© European Union 2014–2025) (https://planetarycomputer.microsoft.com/dataset/sentinel-1-rtc).

The DEM data is produced using Copernicus WorldDEM-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA; all rights reserved.

Annotations were sourced from the Copernicus Emergency Management Service (© European Union, 2012–2025), available at https://emergency.copernicus.eu/.