Binary change detection tells you something changed. Directional multi-label detection tells you what happened. Here's why that distinction matters for urban monitoring and damage assessment.
Most building change detection models output a binary mask. You get a blob of pixels telling you something changed, but not what happened.
For research benchmarks, this is fine. In the real world, "something changed" is only half the story.
The Problem with Binary
A binary mask merges fundamentally different events. A newly built apartment and a demolished warehouse both appear as identical white blobs. The model cannot distinguish them.
Downstream, you can't answer basic operational questions:
- How much new area was built?
- How much area was lost to disaster?
- What is the ratio of growth to demolition?
Binary output only quantifies total changing area. It discards the directional context that makes the data actionable.
The Multi-Label Solution
MBCTD replaces the binary mask with three independent semantic classes: unchanged, newly built, and demolished.
This simple architectural shift unlocks an entirely different level of geospatial analysis.
Why Not Just Diff Footprints?
A common objection: why not extract building footprints before and after, then compare them geometrically?
This fails completely in replacement scenarios, where a building is demolished and rebuilt on the same spot.
If the old and new structures share a footprint, a geometric diff sees zero change. The polygons overlap perfectly, hiding the redevelopment.
MBCTD detects semantic change directly from image pairs. It captures visual shifts in materials and structure that geometry ignores.
Operational Applications
Urban Monitoring
City-scale monitoring now yields concrete metrics. You can compute exact areas of construction versus demolition per neighborhood or district.
Planners can use these metrics to track growth patterns, identify urban decline, and allocate infrastructure budgets with precision.
Damage Assessment
In disasters like wildfires or earthquakes, directional labeling is critical. You don't just need to know what was destroyed—you must know what survived.
MBCTD's "unchanged" class explicitly maps intact buildings. During the LA wildfire assessment, this mapped destroyed and surviving structures together.
For response teams, finding intact buildings dictates where people can shelter and where recovery efforts should focus.
Conclusion
A white blob on a black background is not an insight. To be useful, models must understand the direction of change—separating construction from demolition, and identifying intact structures. This demands a multi-label approach.
Interested in using MBCTD for urban monitoring, damage assessment, or other change detection workflows? Contact me to discuss licensing.
- The MBCTD model: MBCTD Release Post
- The training data: FOTBCD Dataset
- Real-world test: Zero-Shot Damage Assessment on the 2025 LA Wildfires