Transformer model pairs pre- and post-disaster satellite images to detect building damage

Building detection and damage assessment from satellite imagery has clear applications in disaster response and urban planning. However, most existing systems rely on single pre-disaster images, operating under the assumption that post-disaster photos—cluttered with destruction—would confuse the model.

The dual-image approach#

A new model called SiamixFormer rethinks this strategy by processing both pre- and post-disaster satellite images as a pair. The architecture uses a siamese network—meaning it runs two identical encoders in parallel, one for each temporal snapshot. Rather than treating the images as separate inputs, the system fuses their features together intelligently.

How temporal fusion works#

The key innovation lies in how the model combines information across time. At each stage of the two encoders, outputs feed into a temporal transformer. This component generates queries from the pre-disaster image while drawing keys and values from the post-disaster image. This cross-temporal attention mechanism allows the model to learn which pixels changed and what changed them—directly incorporating time as a dimension of the analysis.

By keeping transformer encoders' large receptive fields intact through this fusion approach, the system preserves spatial context better than traditional CNN-based methods would.

Validation and results#

The researchers tested SiamixFormer on multiple benchmark datasets: xBD and WHU for pure building detection tasks, and LEVIR-CD and CDD for detecting changes between image pairs. The model outperformed existing state-of-the-art approaches on all evaluated datasets.