Add Google Earth Engine accessor for satellite imagery (#53) - #76

Thanks for this contribution @KarnakOza — the accessor is well-structured and closely follows the conventions of copernicus_cds.py (optional-import guard, ClassVar metadata, clear ImportError/RuntimeError messages). The index math and the no-imagery ValueError are good. However, there are two blocking issues that contradict the PR's own description, plus some correctness and test-hygiene concerns.

Blocking

1. Entry point was never registered — The PR description promises "add the entry point so the accessor is discoverable via the standard source registry", but pyproject.toml only adds the optional extra. There is no line under [project.entry-points."epidatasets.sources"] (pyproject.toml:136). Without it:

• epidatasets.get_source("google_earth_engine") raises KeyError,
• the source won't appear in list_sources(),
• _registry.py:24 (_load_entry_points) won't find it.

The README still says "33 registered (via plugin registry)" — and after this PR that number is still correct, which is exactly the symptom of the missing registration.

➕ Please add, e.g.:

google_earth_engine = "epidatasets.sources.google_earth_engine:GoogleEarthEngineAccessor"

2. Not added to sources/__init__.py — Every other accessor is listed in src/epidatasets/sources/__init__.py:14-47 (under TYPE_CHECKING) and in __all__. GEE is missing from both, breaking the convention every other source follows. Please add it to both sections.

Correctness

3. Landsat C2 L2 surface-reflectance scaling is not applied (google_earth_engine.py:181). SR_B4/5/6 are raw DNs; reflectance = DN * 0.0000275 − 0.2. For normalized-difference ratios the multiplicative factor cancels, but the additive offset (−0.2) does not, biasing every index. Please either apply image.multiply(0.0000275).add(-0.2) or document explicitly why the offset is ignored.

4. Sentinel-2 / MODIS are advertised but unused — COLLECTIONS declares sentinel2_sr and modis_vi, the docstring lists them, but every method hardcodes Landsat 8. In tropical epidemiology (the stated use case), Landsat's 16-day revisit + the 20% cloud filter means the ValueError in _landsat_composite will fire frequently. Consider a sensor="landsat8"|"sentinel2" parameter, or at least drop the unused entries to avoid false advertising.

Test hygiene

5. TestGoogleEarthEngine.test_initialization will ERROR (not skip) on any machine without EE auth (test_accessors.py:893). The accessor fixture instantiates the class, which calls ee.Initialize() in __init__. Since test_initialization lacks @requires_external_api, CI without earthengine-api installed/authenticated will report an error during setup rather than a skip. Options:

• mark test_initialization with @requires_external_api too, or
• add pytest.importorskip("ee") in the fixture, or
• split into a mocked unit test + a live external test (best — matches the responses-based pattern used by TestDiseaseSh).

6. test_no_imagery_raises_clear_error is flaky (test_accessors.py:937) — it assumes a 1-day Amazon window has zero <20%-cloud Landsat scenes. A single cloud-free acquisition on 2021-03-01 would flip this. Better to monkeypatch the collection size.

Minor / nits

• 7. logging.basicConfig(level=logging.INFO) at module import (google_earth_engine.py:43) reconfigures the root logger on every import. (Same anti-pattern exists in copernicus_cds.py:52, so it's consistent — but worth fixing repo-wide eventually.)
• 8. _reduce_mean silently returns None on a bad band name rather than raising (google_earth_engine.py:188).
• 9. earthengine is intentionally excluded from the all extra (pyproject.toml:105) — reasonable given interactive auth, just confirming it's deliberate.
• 10. Modern str | None hints here vs. Optional[str] in copernicus_cds.py — minor style drift; the new style is preferable.
• 11. README's "Optional extras: 11" update is correct, but the "Data sources: 33 registered" line was not touched (already stale — the entry-point list has 34 — but this PR doesn't fix it).

Suggested changes before merge

• Add entry point to pyproject.toml (item 1)
• Register in sources/__init__.py (item 2)
• Fix test_initialization so it skips cleanly without EE (item 5)
• Apply Landsat SR scaling or document the omission (item 3)

The accessor code itself is solid; once these are addressed the PR will deliver what its description claims and keep CI green on machines without GEE credentials.