Software release note

Introduction

This section constitutes the Software release note (SRN) for eo-data-embedding, release v0.1.0 (the package __version__ is 0.1.0). This is the initial public release.

Software release overview

This Software release note contains release information for the eo-data-embedding, providing:

• version of the release,

• an overview of the contents of the release,

• status of SPRs, SCRs and SW&D related to the release, and

• advice for use of the release.

Version: v0.1.0. This release corresponds to the v0.1.0 tag on the EOPF repository; subsequent releases are tracked through GitHub/EOPF releases.

Status of the software

Evolution since previous version

eo-data-embedding is a multi-modal geospatial embedding toolkit built on a frozen Vision Transformer foundation model (Clay v1.5). As the initial release, v0.1.0 has no previous version. It delivers the following capabilities:

• Similarity search — FAISS retrieval over frozen Clay embeddings, with mAP / precision@k metrics.

• Few-shot classification — a linear probe on frozen embeddings demonstrating the foundation-model label-efficiency benefit against a from-scratch CNN baseline.

• Bitemporal change detection — a supervised Δembedding change probe on OSCD (the zero-training distance baseline is reported but rejected, see limitations).

• Plug-and-play CPU demo — a Gradio UI requiring no GPU and no model at runtime.

The toolkit is config-driven (configs/default.yaml) and exposed through the eo-data-embedding CLI. CI runs lint, unit tests and a CPU Docker image build.

Known problems or limitations

The following limitations are reported honestly per the project V&V honesty practice:

• Change-detection seasonality limit. Zero-training two-date embedding distance is rejected by experiment (ROC-AUC at or below chance) because of a seasonality/phenology confound: unchanged vegetated tiles often move more in embedding space than truly changed urban tiles. The supervised Δembedding probe only partially recovers change (F1 0.510, Kappa 0.231, ROC-AUC 0.640); the phenological layer is an open limitation, with time-series modelling scoped as follow-on work.

• CPU smoke backbone vs Clay. The synthetic CPU smoke / sanity path uses a small timm ViT (random weights), not the full Clay v1.5 model, so it verifies pipeline integrity rather than representation quality. Embedding extraction with Clay needs a GPU and the Clay checkpoint.

• Coverage gate open (REQ-N-05). Test coverage is currently unmeasured; pytest-cov and a CI threshold are not yet wired in.

• Dependency security note. eopf == 2.8.1 hard-pins a vulnerable starlette; the fix must come from a newer eopf (tracked, reported by Trivy as allow_failure).

Unsolved SPRs and approved SW&Ds for this version are tracked via the corresponding issues on the EOPF/GitHub project.

Advice for use of the software configuration item

• Use the CPU demo / app path for plug-and-play evaluation; it requires no GPU, dataset or model at runtime.

• The phase subcommands (extract, search, probe, change) require a source checkout (git clone + pip install -e .); extract and change additionally require a GPU and the Clay checkpoint.

• Run eo-data-embedding smoke on a fresh environment as a green-light gate before a full run.

• Requires Python 3.11. Install CPU-matched torch/torchvision wheels to avoid a torchvision::nms ABI mismatch.

On-going changes

Planned evolution, tracked via issues on the EOPF/GitHub project:

• Close the REQ-F-05 phenological-layer limitation via time-series change detection (e.g. SpaceNet 7 / DynamicEarthNet).

• Wire pytest-cov and a ≥ 70 % coverage gate into CI to close REQ-N-05.

• Adopt a newer eopf once available to clear the pinned starlette vulnerability.

Future releases and their change history are tracked through GitHub/EOPF releases (there is no standalone CHANGELOG in this release).