A research model called SiamixFormer beats existing benchmarks by feeding disaster imagery in pairs instead of one frame at a time.
Most building detection models rely only on pre-disaster images, on the reasoning that post-disaster rubble confuses classifiers. SiamixFormer challenges that assumption directly. The model uses a Siamese architecture — two parallel encoders, one for each image — and routes their outputs through a temporal transformer that treats pre-disaster features as queries and post-disaster features as keys and values. That asymmetric fusion lets the model reason about what changed rather than just what exists. A lightweight MLP decoder sits at each stage, keeping the design from ballooning in complexity.
The approach matters because damage assessment after earthquakes or floods currently requires significant manual overhead. A model that can flag destroyed or altered structures automatically from satellite pairs could meaningfully accelerate triage for rescue teams and urban planners. The temporal transformer also preserves large receptive fields that convolutional networks tend to compress away — a practical advantage when scanning wide areas.
SiamixFormer was evaluated on xBD and WHU for building detection and on LEVIR-CD and CDD for change detection, outperforming state-of-the-art models on each respective task. The architecture is not production software yet — it is a research paper — so the gap between benchmark numbers and operational deployment remains the usual caveat.