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

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.

💡 Research Summary

The paper introduces Skjold‑DiT, a diffusion‑transformer based foundation model that jointly predicts building‑level climate risk and transportation‑network accessibility, targeting intelligent‑vehicle routing and emergency‑response planning. Existing urban forecasting systems either focus solely on mobility (e.g., UrbanDiT) or on physics‑based flood modeling, but they lack integrated multi‑modal data, bidirectional housing‑transport coupling, and counterfactual policy simulation. Skjold‑DiT addresses these gaps through three novel components.

1. Norrland‑Fusion: A cross‑modal attention architecture that ingests six data modalities—geospatial coordinates, structural attributes, demographics, infrastructure connectivity, historical climate exposure, and transportation metrics—alongside a graph representation of the road network and service layer (hospitals, shelters). The graph’s topology and edge weights (free‑flow travel time, hazard‑conditioned penalties) are directly embedded into the attention mechanism, allowing the model to capture how road closures or capacity upgrades affect risk estimates.

2. Fjell‑Prompt: A prompt‑based conditioning interface that encodes hazard type, forecast horizon, and intervention description as textual tokens. These tokens are concatenated with building‑level features, enabling zero‑shot transfer to unseen cities with only minimal meta‑data (average climate indices, road density). This design eliminates the need for city‑specific fine‑tuning and supports compositional generalization across hazard‑policy combinations.

3. Valkyrie‑Forecast: A counterfactual simulator that applies deterministic feature‑edit functions Φ _P to the input features based on a policy prompt P (e.g., green infrastructure, retrofits, road capacity upgrades). The edited features X′ are fed into a conditional diffusion process pθ(Y | X′, G), producing probabilistic trajectories of risk indicators (flood depth, heat stress, structural vulnerability) and derived accessibility scores (reachability, travel‑time inflation, route redundancy). Although the “do(P)” notation does not claim formal causal identifiability, it provides a practical framework for “what‑if” analysis.

The authors release the Baltic‑Caspian Urban Resilience (BCUR) dataset, comprising 847,392 building observations across six cities (including Copenhagen and Baku). Each building is annotated with multi‑hazard labels (flood depth, heat‑stress index) and transportation accessibility metrics, and is linked to high‑resolution satellite/air imagery, elevation maps, and a road‑graph enriched with emergency‑service nodes.

Experimental evaluation covers four axes: (i) risk prediction accuracy (RMSE, macro‑F1), (ii) calibration quality (expected calibration error), (iii) transportation‑relevant outcomes (emergency vehicle reachability, travel‑time inflation), and (iv) policy‑scenario impact. Baselines include GNN, ST‑GCN, the original UrbanDiT, and a physics‑based flood model. Skjold‑DiT outperforms baselines by ~12 % in risk prediction, reduces calibration error by 30 %, and improves emergency‑vehicle reachability by an average of 8 % under simulated interventions. Moreover, it produces calibrated, uncertainty‑aware accessibility layers that can be directly consumed by intelligent‑vehicle routing engines.

Limitations are acknowledged: the model does not incorporate long‑term (≥10 year) climate dynamics with physical constraints, making extreme‑event extrapolation uncertain; counterfactual simulations rely on deterministic feature edits, which may not capture complex interactions among simultaneous policies; and the need for building‑level labeled data remains a bottleneck for low‑resource regions. Future work is proposed to integrate physics‑informed neural operators, adopt Bayesian causal networks for stronger identifiability, explore weak‑supervision or self‑supervised pretraining to reduce labeling costs, and develop multi‑policy optimization techniques.

Overall, Skjold‑DiT represents a significant step toward unified, city‑scale, climate‑resilient urban intelligence, bridging the gap between climate risk assessment and intelligent transportation systems, and offering a practical tool for policymakers and autonomous‑vehicle platforms to anticipate and mitigate disaster impacts.