Granite Geospatial

license

Apache-2.0(OSI)

weights

open(Apache-2.0 ungated checkpoint files)

data

documented-not-released(HLS L30 and GEDI L4A sources and the leaf-on compositing method described, assembled training set not published)

code

partial(getting-started and few-shot notebooks over TerraTorch, no pretraining or fine-tuning pipeline)

The checkpoint and its configuration are published under Apache-2.0 and the card names its NASA source products in detail, but the model repository is archived notebooks and the assembled corpus never shipped.

• https://huggingface.co/ibm-granite/granite-geospatial-biomass/raw/main/README.md recorded 2026-09-10

  Apache-2.0 card stating that the weights and the configuration file that defines them are made available, naming HLS L30 and GEDI L4A as the training data with the cloud-free compositing method, and pointing to getting-started notebooks rather than training code.

• https://huggingface.co/api/models/ibm-granite/granite-geospatial-biomass recorded 2026-09-10

  gated: false; private: false; license apache-2.0; biomass_model.ckpt, biomass_model_lightning.ckpt, and config.yaml present.

• https://raw.githubusercontent.com/ibm-granite/granite-geospatial-biomass/main/README.md recorded 2026-09-10

  Describes the repository as public notebooks and utilities for working with the fine-tuned model, with no training or fine-tuning pipeline.

127 downloads in the trailing 30 days for the declared biomass checkpoint.

• https://huggingface.co/api/models/ibm-granite/granite-geospatial-biomass recorded 2026-09-10

  downloads: 127; gated: false; private: false.

Only the adaptation pathway is stated directly. The multi-instrument decline turns on which HLS member the card names: the declared input is HLS L30, the Landsat-only product, not the S30 Sentinel-2 member, so it is one instrument rather than the two the HLS name suggests. Confidence is medium because the card contradicts itself on that point — the text says L30 while its own Training Data link points at the S30 product page — though both readings leave a single input product. The GEDI lidar is a label source rather than an input, masked-reconstruction pretraining does not count as generative, and no second scale is specified.

• https://huggingface.co/ibm-granite/granite-geospatial-biomass/raw/main/README.md recorded 2026-09-10

  Documents TerraTorch inference from a released config plus notebooks walking zero-shot and few-shot fine-tuning, with HLS L30 optical imagery as the only model input and above-ground biomass regression as the only output.