Urban Spatio-Temporal Foundation Models for Climate-Resilient Housing: Scaling Diffusion Transformers for Disaster Risk Prediction

TL;DR

Skjold-DiT introduces a diffusion-transformer-based framework that fuses multi-modal urban data with transportation-network signals to forecast building-level climate risk and generate hazard-conditioned accessibility layers for intelligent-vehicle routing. The architecture combines Norrland-Fusion for multi-modal encoding, Fjell-Prompt for cross-city zero-shot transfer, and Valkyrie-Forecast for probabilistic counterfactual simulation, validated on the BCUR dataset of 847,392 buildings across six cities. Results show strong predictive performance, robust cross-city generalization, and calibrated uncertainty, enabling scenario analysis for emergency routing and equity-focused policy. The work demonstrates practical deployment pathways with edge-cloud pipelines and policy-relevant insights aligned with World Urban Forum 13 resilience priorities, offering a scalable tool for climate-resilient urban planning and transportation optimization. \\Delta t$-dependent forecasts, hazard-conditioned routing, and explicit intervention prompts position diffusion-transformer urban foundation models as a principled approach for integrating climate science, housing vulnerability, and mobility in smart-city decision support.

Abstract

Climate hazards increasingly disrupt urban transportation and emergency-response operations by damaging housing stock, degrading infrastructure, and reducing network accessibility. This paper presents Skjold-DiT, a diffusion-transformer framework that integrates heterogeneous spatio-temporal urban data to forecast building-level climate-risk indicators while explicitly incorporating transportation-network structure and accessibility signals relevant to intelligent vehicles (e.g., emergency reachability and evacuation-route constraints). Concretely, Skjold-DiT enables hazard-conditioned routing constraints by producing calibrated, uncertainty-aware accessibility layers (reachability, travel-time inflation, and route redundancy) that can be consumed by intelligent-vehicle routing and emergency dispatch systems. Skjold-DiT combines: (1) Fjell-Prompt, a prompt-based conditioning interface designed to support cross-city transfer; (2) Norrland-Fusion, a cross-modal attention mechanism unifying hazard maps/imagery, building attributes, demographics, and transportation infrastructure into a shared latent representation; and (3) Valkyrie-Forecast, a counterfactual simulator for generating probabilistic risk trajectories under intervention prompts. We introduce the Baltic-Caspian Urban Resilience (BCUR) dataset with 847,392 building-level observations across six cities, including multi-hazard annotations (e.g., flood and heat indicators) and transportation accessibility features. Experiments evaluate prediction quality, cross-city generalization, calibration, and downstream transportation-relevant outcomes, including reachability and hazard-conditioned travel times under counterfactual interventions.

Figures (5)

• Figure 1: Baltic--Caspian Urban Resilience (BCUR) dataset: cities, modalities, and annotation types.
• Figure 2: Overview of the proposed Skjold-DiT framework for transportation-aware climate-resilient housing risk prediction.
• Figure 3: Long-term forecast performance comparison under the temporal split. Error bars represent 90% credible intervals. Reported values are averaged over 5 random seeds.
• Figure 4: Flood vulnerability by income quintile in Copenhagen. Without intervention (red bars), low-income residents face 2.8× higher 10-year flood risk compared to high-income residents. Targeted green infrastructure policies (green bars) reduce the disparity to 1.4×, demonstrating the importance of equity-centered climate adaptation strategies.
• Figure 5: Reliability diagram for 10-year flood predictions showing near-perfect calibration. Blue bars represent Skjold-DiT predictions across different confidence bins, while the black diagonal line indicates perfect calibration. Red dotted lines show calibration gaps. The Expected Calibration Error (ECE) of 0.037 demonstrates highly reliable uncertainty estimates essential for risk-sensitive policy applications.