TerraMind-NYC-Adapters

LoRA adapters that specialise ibm-esa-geospatial/TerraMind-1.0-base (IBM-ESA TerraMind 1.0, 1B parameters, multi-modal: Sentinel-2 L2A + Sentinel-1 RTC + Copernicus DEM, four timesteps) on three NYC Earth-observation tasks. Trained on AMD Instinct MI300X via AMD Developer Cloud. Apache-2.0.

GitHub mirror of the model on Hugging Face: huggingface.co/msradam/TerraMind-NYC-Adapters.

Adapters in this family

Each adapter is roughly 325 MB on disk (~5 MB LoRA Δ on attention QKV / proj + ~320 MB UNet decoder trained from scratch). The 1.45 GB TerraMind base sits on disk once and is shared across all adapters.

Demo segmentations

Buildings adapter

Manhattan midtown — model finds essentially every building:

Jamaica Bay — model correctly finds 0.18 % buildings:

Central Park — mixed urban / vegetation:

LULC adapter (5 classes: water / impervious / vegetation / bare / building)

Manhattan midtown — dominantly impervious:

Jamaica Bay — 96 % water:

Central Park — vegetation visible:

Sniff-test results

Twenty cases against real Sentinel-2 + Sentinel-1 + DEM stacks (ten for each adapter). All twenty pass.

Buildings adapter

LULC adapter

Threshold-sweep operating points (buildings)

Recommended operating points: 0.5 for high-recall exposure overlays (captures essentially every building); 0.6 for higher precision. Above 0.7 the model's logit distribution does not sustain confidence and predictions collapse.

Benchmark (M3 Air, CPU fp32)

Install and use

git clone https://github.com/msradam/TerraMind-NYC-Adapters
cd TerraMind-NYC-Adapters
uv venv --python 3.12
uv pip install -e ".[dev]"

Direct usage (downloads 1.45 GB TerraMind base + 305 MB adapter on first run):

from terramind_nyc_adapters import load_terramind_adapter

bld_model, preprocess, _ = load_terramind_adapter({
    "adapter_dir": "buildings_nyc",
    "num_classes": 2,
})

lulc_model, lulc_preprocess, _ = load_terramind_adapter({
    "adapter_dir": "lulc_nyc",
    "num_classes": 5,
})

Training

Full training methodology is in docs/TRAINING.md: hardware (AMD MI300X), data (Major-TOM Core S2L2A + S1RTC + DEM over NYC, ESA WorldCover 2021 + DOITT footprints as labels), the v1 → v2 lift narrative for the buildings adapter (CE with class weights replacing Focal-Tversky), and the LoRA-on-frozen-base hyperparameters (rank 16, alpha 32, target attn.qkv and attn.proj across 24 transformer blocks).

Where this fits

One of three NYC fine-tuned foundation models in this family.

• Reproduction harness, Streamlit demo, and probe tooling: github.com/msradam/riprap-models.
• Sister repos: Granite-TTM-r2-Battery-Surge and Prithvi-EO-2.0-NYC-Pluvial.
• Parent system: Riprap-NYC.

Sources

• Sentinel-2 / Sentinel-1 imagery via Microsoft Planetary Computer (Copernicus Open Data License).
• NYC DOITT building footprints: NYC OpenData public domain (5zhs-2jue).
• ESA WorldCover 2021 v200 under the ESA CCI Open Data Policy (CC-BY-4.0).

AI-assisted authoring

Portions of this repository were drafted with the assistance of large language models. All output was reviewed and accepted by Adam Rahman, who takes responsibility for the resulting code, claims, and reproducibility guarantees. The full disclosure is in NOTICE.

License

Apache-2.0. See LICENSE.