Parameter Guide

This page provides practical guidance for the most important brdr parameters.

relevant_distance

  • Meaning: maximum spatial search/tolerance distance for alignment behavior.
  • Typical approach: evaluate a distance range, then use predict(...) or evaluate(...).
  • Start small and increase gradually.

profile_performance

  • Location: AlignerConfig(profile_performance=...)
  • Meaning: enable lightweight internal timing during process(...).
  • Output: timings are available in AlignerResult.metadata["performance"].
  • Default: False.
  • Recommendation: enable for tuning/benchmarking runs, disable for normal production runs.

correction_distance

  • Meaning: technical cleanup tolerance used in post-processing.
  • Keep this small relative to your data resolution.
  • If too large, small valid features may be over-smoothed.

mitre_limit

  • Meaning: controls corner behavior in buffer operations.
  • Higher values keep sharper corners; lower values reduce spikes.

snap_strategy

  • Meaning: controls vertex/edge preference in snapping workflows.
  • Use strict vertex-based modes when reference vertices are authoritative.
  • Use no-preference mode when edge proximity is equally valid.

od_strategy (Open Domain)

  • Meaning: defines behavior where no polygon reference coverage exists.
  • Practical note: when the reference dataset has only lines/points and no polygonal coverage, open-domain strategy effects are intentionally limited.
  • Choose explicitly based on business semantics:
    • keep as-is
    • exclude
    • snap to virtual/reference boundaries

Tuning workflow

  1. Fix a representative sample dataset.
  2. Sweep relevant_distance.
  3. Enable profile_performance and inspect metadata["performance"].
  4. Inspect stability and geometric deltas.
  5. Lock od_strategy and snap_strategy.
  6. Validate on a larger holdout set.

Performance notes

  • brdr reuses reference candidates per thematic geometry across multiple relevant distances.
  • evaluate(...) includes caching for repeated observation comparisons on identical predicted geometries.
  • These optimizations are transparent; no extra configuration is needed besides optional profiling.