Detection logic

The code searches for persistent departures from a seasonal baseline, not simple before/after image differences.

flowchart LR
  A[History scenes] --> B[Fit harmonic seasonal baseline]
  C[Monitoring scenes] --> D[Predict expected S1 signal]
  B --> D
  D --> E[Residuals and z-scores]
  E --> F[VV/VH fusion]
  F --> G[Persistence rule]
  G --> H[Durability filter]
  H --> I[Flag + change date]

Improvement in senbygg_improved.py	Why it matters
Increase-only directional scoring	Construction usually raises radar response; ignoring decreases removes many vegetation/snow/moisture false alarms.
Two-harmonic seasonality	Better represents high-latitude freeze/thaw seasonality.
Persistence	Requires multiple consecutive scenes above threshold.
Summer durability	Checks that a trigger is not a transient spring/snow signal.
Candidate polygon aggregation	Avoids noisy pixel-by-pixel operational decisions.

Recommended defaults: z_threshold=2.0, persistence_n=3, durability_n=2, winter_months={12,1,2}, poly_agg="p75".