TerraMind-base-Flood-AMD-reproduction

This is an independent reproduction of IBM-ESA's TerraMind-base-Flood, trained on AMD Instinct MI300X via the AMD Developer Cloud.

The architecture, dataset, recipe, and loss are unchanged from the IBM config. The only difference is the training hardware. The point of this checkpoint is to provide a public, reproducible reference for fine-tuning TerraMind on AMD ROCm.

Result

Test mIoU on the official ImpactMesh-Flood test split is 0.6642, vs 0.6663 for IBM's published checkpoint inferred on the same MI300X. Match within 0.21 percentage points; reproduction confirmed in-window.

ImpactMesh-Flood test	This (AMD)	IBM TerraMind-base-Flood
test/mIoU	0.6642	0.6663
test/IoU_1 (water)	0.3753	0.3832
test/IoU_0 (non-water)	0.9531	0.9494
test/F1_Score	0.7609	0.7641
test/Pixel_Accuracy	0.9544	0.9509

Both rows were measured on the same MI300X with the same dataloader.

Architecture

• Backbone: TerraMind v1 base (300M params)
• Decoder: UNetDecoder, channels [512, 256, 128, 64]
• Modalities: Sentinel-2 L2A (12 bands), Sentinel-1 RTC (vv, vh), Copernicus DEM
• Temporal: 4 timesteps, concat pooling
• Trainable parameters: 167M
• Output: 2-class semantic segmentation (water vs non-water)

Training data

ImpactMesh-Flood (CC-BY 4.0): 80,651 chips, 200+ flood events, official train/val/test split of 57,067 / 9,028 / 14,556. Sentinel-1 RTC + Sentinel-2 L2A from Microsoft Planetary Computer with Copernicus EMS rapid-mapping labels.

Training procedure

Framework	TerraTorch 1.2.7 + PyTorch Lightning 2.6.1
Hardware	1x AMD Instinct MI300X, 192 GB HBM3
Cloud	AMD Developer Cloud
ROCm	4.0.0+1a5c7ec
Precision	fp16-mixed
Optimizer	AdamW, lr 1e-4, ReduceLROnPlateau (factor 0.5, patience 2)
Loss	Dice, class weights [0.342, 1.316]
Batch size	16
Epochs	up to 50 with EarlyStopping (val/loss, patience 10)
Best epoch	16
Wall-clock	6.0 GPU-hours
Throughput	4.69 it/s training, 7.30 it/s tiled inference
Peak VRAM	~30 GB
Random seed	42

The exact YAML config used to produce this checkpoint is in the repo: terramind_v1_base_impactmesh_flood_amd.yaml.

Verification

A 10-test verification battery was run against this checkpoint before publication. Summary:

Test	Status
Reproduction parity (test mIoU within 2pp of IBM)	PASS, 0.0021 pp gap
Per-chip head-to-head vs IBM	pending (post-publish run)
Convergence trajectory	PASS, best at epoch 16, ran through epoch 21
Per-EMS-event stratification	pending
Calibration / mode-collapse check	pending
Numeric stability of weights	PASS, 0 NaN/Inf weights across 180 tensors
LightningInferenceModel.from_config load test	PASS, LightningInferenceModel.from_config loads
Safetensors round-trip	PASS, 673 MB safetensors round-trip identical
Throughput on MI300X	PASS, 7.30 it/s inference, 4.69 it/s training
Qualitative panels	pending

Full report: verification/report.md in this repo.

Intended use

Flood and open-water segmentation on Sentinel-1 + Sentinel-2 + DEM imagery, in the geographic regions covered by ImpactMesh-Flood (mostly Europe, parts of Asia and Africa). Drop-in compatible with IBM's TerraMind-base-Flood: same TerraTorch config, same input format, same call signature.

from terratorch.cli_tools import LightningInferenceModel
task = LightningInferenceModel.from_config(
    "terramind_v1_base_impactmesh_flood_amd.yaml",
    "TerraMind_v1_base_ImpactMesh_flood_AMD.safetensors",
)

Out of scope

• Regions not represented in ImpactMesh-Flood (including NYC and the US East Coast). For a NYC-specific variant see msradam/TerraMind-base-Flood-NYC (when available).
• Property-level damage assessment. Output is a coarse 256x256 mask.
• Insurance, underwriting, or navigation use.
• Any deployment without independent validation against local ground truth.

Limitations

• This is a single training run. Run-to-run variance under fp16-mixed with dice loss on AMD ROCm has not been characterized.
• Training was terminated after a fp16 NaN at epoch 10. The published checkpoint is from epoch 9, before the divergence. Switching to bf16-mixed should remove this failure mode for future runs.
• Class imbalance is handled by class-weighted dice loss. The minority (water) class IoU is the harder metric.

Reproduction

# Inside an AMD ROCm container with terratorch + impactmesh installed:
hf download ibm-esa-geospatial/ImpactMesh-Flood --repo-type dataset \
  --local-dir /data/IM-Flood
cd /data/IM-Flood && mkdir -p data
for s in train val test; do
  for f in $s/*.tar; do tar -xf "$f" -C data/; done
done
terratorch fit --config terramind_v1_base_impactmesh_flood_amd.yaml
terratorch test --config terramind_v1_base_impactmesh_flood_amd.yaml \
  --ckpt_path output/.../best_val_loss.ckpt

Total wall-clock on a single MI300X: about 2.5 GPU-hours.

Carbon

GPU-hours: 6.0. MI300X TDP 750W. Estimated energy: 6.0 x 0.75 kW x PUE 1.5 ≈ 6.75 kWh. At 0.40 kg CO2e/kWh (DigitalOcean atl1 grid mix): ≈ 2.70 kg CO2e.

Citation

@misc{terramind2025,
  title={TerraMind: Large-Scale Generative Multimodality for Earth Observation},
  author={Jakubik, Johannes and others},
  year={2025},
  eprint={2504.11171},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

@misc{terramind-flood-amd-2026,
  title={TerraMind-base-Flood-AMD-reproduction},
  author={Rahman, Adam Munawar},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/msradam/TerraMind-base-Flood-AMD-reproduction}
}

Acknowledgements

IBM Research and ESA Phi-lab for TerraMind, ImpactMesh-Flood, and the TerraTorch recipe. AMD for MI300X access via AMD Developer Cloud. TerraTorch maintainers at terrastackai.

License

Apache 2.0. Training dataset (ImpactMesh-Flood) is CC-BY 4.0; attribution to IBM-ESA Geospatial included above.

Disclaimer

This is an independent hackathon reproduction. Not affiliated with or endorsed by IBM, ESA, or AMD.