Key findingsΒΆ
- Directional increase-only scoring sharply reduced raw false positives because construction is expected to raise radar response.
- Candidate/polygon workflows avoided the whole-AOI base-rate problem and produced a reviewable queue.
- Durability filtering helped remove snow/vegetation/moisture transients.
- Coherence fusion was the strongest optional recall lever in current validation.
- Sentinel-2 fusion helped recover vegetation-to-built changes but remains cloud/season dependent.
For small Norwegian buildings near the Sentinel-1 GRD pixel scale, intensity-only recall can plateau around 0.3-0.5. That is a sensor/signal limit as much as an algorithm limit.