ViTCapsNets: hybrid vision transformers and capsule networks architecture for automated building damage assessment using multitemporal satellite imagery

Rapid and accurate building damage assessment following natural disasters is critical for effective emergency response. Existing deep learning approaches predominantly rely on binary classification or standard convolutional architectures that discard pose and spatial relationship information through max-pooling. This paper presents ViTCapsNets, a family of hybrid architectures integrating Vision Transformer variants with Capsule Networks for four-class ordinal building damage assessment from bi-temporal satellite imagery. Within this family, we evaluate standard Vision Transformer (ViT) and Swin Transformer backbones — the latter a hierarchical extension of ViT replacing global self-attention with shiftedwindow attention — paired with a Capsule Network head employing dynamic routing to preserve equivariant spatial partwhole relationships. Applied to the xBD dataset across six disaster types, the SwinCapsNets variant achieves a Cohen's Kappa of 0.675 and Matthews Correlation Coefficient of 0.676 — metrics introduced in this work to complement the dataset's standard F1-score evaluation — with 42M parameters, outperforming all ViTCapsNets configurations. A multiscale preprocessing strategy across three input resolutions (32×32, 64×64, 128×128 pixels) mitigates the Modifiable Areal Unit Problem inherent to buildings of variable footprint size. Beyond the scientific contribution, we present ViDa Caps — a minimum viable product integrating the best-performing model into a QGIS plugin — as an example of responsible science that closes the loop between research and operational societal impact.