new parameter LCCClassesToReplaceNNMethod; require LandR (>= 1.2.0.9005)

[achubaty]

Exposes the method argument added to LandR::convertUnwantedLCC() in PredictiveEcology/LandR#197 as a module parameter.

Depends on PredictiveEcology/LandR#197 — merge that first.

Targets development (not main, which is 128 commits behind it).

New parameter

LCCClassesToReplaceNNMethod, passed straight through to convertUnwantedLCC(method = ). Both options allocate each LCCClassesToReplaceNN pixel a neighbouring class drawn in proportion to that class's abundance in the pixel's neighbourhood; they differ only in where the draw comes from:

value	behaviour
"nearestRandom" (default)	draws from the RNG — preserves the stochastic allocation this module has always had
"nearestWeighted"	keys the draw on the pixel's ground position: deterministic, needs no set.seed(), stable under Cache(), and a grid-aligned crop of the study area gives the same answer as the full extent

The default deliberately keeps existing behaviour. "nearestWeighted" is there for workflows that need reproducibility without seed management, or that develop on a cropped subset before scaling up.

Why the LandR floor moves to 1.2.0.9005

Two reasons:

1. method doesn't exist before it, so an older LandR fails with a clear unused-argument error rather than silently ignoring the parameter.
2. LandR 1.2.0.9004 dropped the newPossLCC column used below to write replacement classes back into rstLCCAdj. The is.null() guard there silently stopped firing, leaving rstLCCAdj — and therefore ecoregionMap — showing the un-replaced classes. LandR#197 restores it; the comment now records the gap so the floor isn't lowered later without noticing.

Context worth knowing

LandR 1.2.0.9004 briefly replaced this function's stochastic search with a deterministic nearest-class rule that broke distance ties to the lowest land-cover class. Ties turn out to be common — 35–41% of unwanted pixels on real landscapes — and because the Canada LCC codes run non-vegetated → non-forest vegetation → forest, that biased imputation systematically out of forest. Pooled over four real SCANFI+FAO landscapes, versus the previous implementation:

class	cover type	old	1.2.0.9004 rule
50	shrubs	9.63%	16.25% (1.69×)
220	broadleaf	15.12%	8.75% (0.58×)
230	mixedwood	6.97%	1.96% (0.28×)

Non-forest vegetation went 11.3% → 18.6% of replaced pixels — about one in fourteen pixels that would have been forest wasn't. For this module that is not cosmetic: a pixel imputed as shrubs or herbs carries no tree cohorts at all.

LandR#197 removed that rule. Neither value of this parameter reintroduces it — both weight by local abundance and land within 0.9–1.0× of the old algorithm on every cover type.

Testing

• Module metadata parses; SpaDES.core::moduleParams() registers the new parameter and moduleMetadata() reports the bumped floor and version.
• Not yet exercised end-to-end — that needs LandR#197 merged and installed first.

Related

• convertUnwantedLCC(method = 'nearestRandom') under Cache(): one draw replayed across replicates #101 — Cache() does not key on RNG state, so under the "nearestRandom" default a cached call replays a single draw across replicates. Pre-existing: this call has always been stochastic and cached. Filed separately rather than folded in here, since fixing it changes caching behaviour.
• fork default: LCCClassesToReplaceNNMethod = 'nearestWeighted'; require LandR (>= 1.2.0.9005) FOR-CAST/Biomass_borealDataPrep#1 — the fork-side counterpart, already merged, which defaults to "nearestWeighted" instead for the projects consuming that fork.

🤖 Generated with Claude Code