A pixel can describe more than one event
A standard building-change mask answers whether a pixel changed. MBCTD asks what happened there and does not force the answers to be mutually exclusive.
The model produces three full-resolution probability maps:
| Output | Meaning |
|---|---|
UNCHANGED |
A building is present at both dates |
DEMOLISHED |
A building is present before and absent after |
NEW |
A building is absent before and present after |
Because each channel has its own sigmoid probability, the same pixel can be both DEMOLISHED and NEW. That overlap represents a replacement: an earlier building was removed and another occupies the same location. Collapsing the output for display produces a fourth visual class, but the underlying prediction remains three independent masks.
Architecture
Both dates pass through the same ConvNeXt-Base encoder, initialized with DINOv3 LVD-1689M weights. The shared encoder returns features at four scales, from one-quarter to one-thirty-second of the input resolution.
At every scale, MBCTD combines the temporal features as:
[before, after, before - after, |before - after|]A 1 × 1 projection followed by a 3 × 3 convolution turns that four-part representation into a change feature. This retains temporal direction through the signed difference while also exposing the magnitude of change through the absolute difference.
A U-Net-style decoder fuses the four scales and restores the original resolution with learned PixelShuffle upsampling. Encoder skip connections carry semantic and spatial detail; two additional skips derived from the raw image pair recover information at half and full resolution. A final 1 × 1 convolution emits three logits per pixel.
Training data and evaluation
I trained MBCTD exclusively on the larger instance-annotated FOTBCD collection: more than 220,000 georeferenced before-and-after pairs from France, sampled as 256 × 256 patches. The labels provide the unchanged, demolished, and new building polygons needed for dense multi-label supervision.
FOTBCD supports all three output channels, so I evaluated it both as binary change detection and per class. LEVIR-CD+ only supplies binary change ground truth; there, demolished and new predictions are combined into a single changed mask. The decision threshold was selected separately by best F1 on each benchmark.
Binary change results
| Dataset | Threshold | Precision | Recall | F1 | Change IoU | mIoU |
|---|---|---|---|---|---|---|
| FOTBCD | 0.70 | 0.8948 | 0.9201 | 0.9073 | 0.8303 | 0.9094 |
| LEVIR-CD+ | 0.75 | 0.7694 | 0.8137 | 0.7909 | 0.6541 | 0.8180 |
FOTBCD per-class IoU
| Class | IoU |
|---|---|
| Unchanged | 0.7774 |
| Demolished | 0.8166 |
| New | 0.8198 |
The LEVIR-CD+ result tests geographic and dataset transfer, but only at the binary level. It cannot measure whether the model assigned the correct direction of change.
Making the model inspectable
The release includes the model definition, configuration, checkpoint loading, preprocessing, prediction helpers, and an interactive Gradio application. The interface accepts a before-and-after pair, exposes the confidence threshold, and supports two inference modes:
- Full resolution runs the pair at its native size.
- Patch mode tiles large images into 256-pixel windows and stitches their predictions.
It returns the three binary masks, a collapsed display map, an overlay on the later image, and per-class pixel-coverage statistics. The checkpoint and inference code are released under CC BY-NC 4.0.
A zero-shot wildfire stress test
I also ran the released pipeline on temporal imagery from the 2025 Los Angeles wildfires without retraining. The model often responded to complete structural disappearance or replacement, but it was less reliable for partial damage, subtle roof changes, and scenes dominated by vegetation or shadow.
That experiment measures transfer, not disaster-assessment readiness. Burn damage has different visual evidence and operational labels from the French building transitions used for training. The detailed wildfire evaluation documents those failure modes rather than presenting the model as a deployed damage product.
Limitations
MBCTD is building-specific and cannot describe road, vegetation, or land-use change. Its training geography is France, and the lower LEVIR-CD+ score shows that transfer remains imperfect. Sigmoid thresholds are benchmark-dependent and require calibration for a new sensor or region. Overlapping masks can represent replacement, but they do not establish why a building changed. The public repository provides inference and a pretrained checkpoint rather than the complete training pipeline, and any operational use—especially disaster response—requires target-domain labels, uncertainty analysis, and human review.