Kelp AEF

kelp-aef is a Python pipeline for testing whether AlphaEarth/Satellite Embedding features can reproduce Kelpwatch-style kelp canopy labels in a restartable, artifact-producing workflow.

Kelpwatch is a Landsat-derived weak-label product, not independent field truth. Results should be framed as learning or reproducing Kelpwatch-style labels, not as proving true kelp biomass.

Current Scope

The current happy path is the annual-max retained-domain HGB hurdle workflow:

Kelpwatch labels
  + AlphaEarth annual embeddings
  -> Kelpwatch-native 30 m target grid
  -> CRM/domain mask
  -> retained-domain model-input sample
  -> split manifest
  -> HGB binary support model
  -> HGB positive-canopy amount model
  -> expected-value hurdle composition
  -> metrics, maps, visualizers, and model-analysis reports

Phase 4 north-south transfer research is closed. Start with docs/index.md for current routing, closed-phase evidence, and read/skip guidance.

Setup

uv sync

The package installs a kelp-aef CLI:

uv run kelp-aef <command> --config configs/santa_cruz.yaml

Current configs include:

• configs/santa_cruz.yaml for the fast smoke loop.
• configs/central_coast.yaml for the larger Monterey plus Big Sur base run.
• configs/fort_bragg.yaml for the northern Phase 4 contrast region.
• configs/long_beach.yaml for the southern Phase 4 contrast region.
• configs/phase4_pooled_all_regions_annual_max_10pct.yaml for the closed pooled Central Coast plus Fort Bragg plus Long Beach source run.

Data And Artifacts

Keep the git repo focused on source, configs, tests, and docs. Raw data and generated artifacts live outside the repo under:

/Volumes/x10pro/kelp_aef

See docs/data_artifacts.md for the artifact contract and git policy.

Validation

For docs-only changes:

git diff --check

For code, configs, or tests:

make check

Docs

• AGENTS.md: repo instructions for agents.
• docs/index.md: where to start for navigation and current status.
• docs/documentation.md: how to edit docs without duplicating history.
• docs/product.md: product and scientific claim boundary.
• docs/architecture.md: current technical contract.
• docs/todo.md: current queue and closed-phase pointers.
• docs/backlog.md: future idea dump.
• docs/decisions/: accepted decisions and negative/non-promoted experiments.
• docs/report_snapshots/: tracked closed-phase report snapshots.