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 masks
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.
Because of this, you can't answer basic operational questions:
- How much new area was built?
- How much area was lost to a disaster?
- What is the ratio of growth to demolition?
Binary output only quantifies total changing area. It discards the context that makes the data actionable.
The multi-label solution
MBCTD replaces the binary mask with three independent classes: unchanged, newly built, and demolished.
This lets us extract more specific metrics.
Why not just diff footprints?
A common suggestion is to extract building footprints before and after, then compare them geometrically.
This fails 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 a geometric diff ignores.
Operational applications
Urban monitoring
City-scale monitoring can now yield concrete metrics. You can compute exact areas of construction versus demolition per neighborhood or district. Planners can use this data to track growth patterns and allocate infrastructure budgets.
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.
Beyond the binary
Models must understand the direction of change to be truly useful. Separating construction from demolition, and identifying intact structures, requires 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