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

Olaf Yunus Laitinen Imanov; Derya Umut Kulali; Taner Yilmaz

Source-linked AI summary

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

Olaf Yunus Laitinen Imanov, Derya Umut Kulali, Taner Yilmaz

arXiv:2602.06129v1Feb 5, 2026cs.LGcs.AI

TL;DR

Urbanization and climate hazards create a need for quantitative decision support linking housing vulnerability, climate risk, and transportation accessibility. Skjold-DiT addresses this with a multimodal diffusion-transformer framework and reports strong flood prediction, cross-city transfer, and historical validation results, while counterfactual outcomes remain model-based projections.

Problem
Urban resilience requires quantitative, city-scale decision support linking climate risk, housing vulnerability, and transportation accessibility because hazards damage housing and disrupt emergency routing and evacuation planning.
Method
Skjold-DiT integrates heterogeneous urban modalities and transportation-network accessibility through Norrland-Fusion, uses Fjell-Prompt for unseen-city transfer, and applies Valkyrie-Forecast for counterfactual housing-risk trajectories.
Results
94.7% accuracy on 10-year flood-risk classification, 87.2% flood accuracy when transferring zero-shot to Baku, and 85.3% identification of flooded structures in the 2010 Kura River validation are reported.
Takeaways & Limitations
The framework connects vulnerability assessment with transportation access, demographic vulnerability, cross-city transfer, and accessibility-constrained counterfactual analysis for urban resilience planning.
Takeaways & Limitations
Counterfactual outcomes are model-based projections rather than verified realized impacts, and rare compound events can be underestimated.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Urban resilience requires decision support linking climate risk, housing vulnerability, and transportation accessibility. Skjold-DiT addresses gaps in multimodal integration, housing–transportation coupling, cross-city transfer, and counterfactual planning.

Motivation: Climate hazards damage housing and infrastructure while reducing emergency access, evacuation capacity, and transportation-network availability.The paper connects flood, heat, and sea-level hazards to emergency routing, autonomous navigation, traffic management, and urban planning.
Research gap: Few existing frameworks jointly predict housing vulnerability and analyze transportation networks for disaster preparedness and intelligent-vehicle safety.Prior systems generally separate these concerns or focus on traffic forecasting without climate-risk and housing-vulnerability modeling.
Proposed framework: Skjold-DiT links climate-risk prediction with transportation-accessibility signals relevant to intelligent vehicles.Its contributions include multimodal fusion, prompt-based cross-city transfer, and counterfactual intervention simulation.
Related approaches: Existing urban forecasting architectures evolved from recurrent and graph models toward transformers and diffusion transformers, but remain limited in transfer and resilience-aware integration.UrbanDiT supports zero-shot mobility prediction, while this work extends the paradigm to climate-resilient housing and transportation infrastructure.
Related approaches: Physics-based flood simulations can require 10,000+ CPU hours per city-scale scenario, while machine-learning approaches often omit uncertainty or key urban-system dimensions.Reported omissions include infrastructure networks, socioeconomic vulnerability, long-term scenarios, and counterfactual policy evaluation.

C. Diffusion Models for Scientific Applications

The paper frames Skjold-DiT as a multimodal urban foundation model designed to address limitations in existing scientific and urban diffusion applications. Its formulation combines heterogeneous city data with long-horizon prediction, counterfactual generation, and zero-shot transfer.

Diffusion models: Diffusion models provide a scalable generative foundation, while diffusion transformers replace U-Net backbones with transformer architectures.Prior extensions include latent diffusion, classifier-free guidance, and continuous-time score-based formulations.
Research gaps: Existing urban diffusion applications generate traffic or mobility scenarios but lack integrated climate risks, housing vulnerability, transportation signals, and policy counterfactuals.Skjold-DiT is positioned as an extension toward climate-resilient housing and intelligent-vehicle applications.
Research gaps: The framework explicitly includes socioeconomic vulnerability, transportation-housing coupling, and equity-related accessibility considerations.These design goals respond to stated gaps in cross-city deployment, long-term planning, and transportation access disparities.
Problem formulation: The model uses building-level spatial, structural, demographic, infrastructure, climate, and transportation features together with a city graph.The graph includes spatial relationships, infrastructure connectivity, and transportation-network accessibility.
Tasks: Skjold-DiT targets long-horizon risk forecasting, intervention-conditioned alternative futures, and zero-shot generalization to unseen cities.The forecast horizon is ∆t ∈[1, 10] years, and zero-shot transfer uses prompt-based conditioning without fine-tuning.

1) Outputs and Task Definitions:

Skjold-DiT treats building-level risk as a multitask target spanning hazards, structural vulnerability, and transportation accessibility. It supports predictive forecasting and scenario-conditioned generation under documented intervention edits.

Target outputs: The multitask target includes flood depth, heat stress, structural vulnerability, and transportation accessibility scores.Accessibility examples include emergency reachability under hazard-induced road constraints.
Predictive task: The predictive task conditionally generates future building-risk outputs from current multimodal observations and transportation-accessibility metrics.The implementation may use continuous target generation or diffusion-based latent generation with task-specific heads.
Counterfactual task: The counterfactual task generates alternative futures under prompts for green infrastructure, building retrofits, or transportation-network improvements.The notation do(P) describes scenario-conditioned generation rather than formal causal identification from observational data.
Intervention prompts: Counterfactual inputs are produced by a documented deterministic edit map X′ = ΦP(X).The paper specifies edits for drainage, structural scores, damage probability, evacuation-edge capacity, and hazard-conditioned road weights.

3) Transportation Graph and Accessibility Signals:

The transportation component represents roads, services, and hazard-conditioned edge availability as a graph-based accessibility system. Evaluation measures reachability, travel time, route redundancy, predictive quality, risk sensitivity, calibration, and temporal or spatial generalization.

Graph layers: The transportation graph combines physical road infrastructure, emergency facilities and shelters, and hazard-conditioned edge availability.Hazards can remove edges or inflate their weights, enabling accessibility features conditioned on network disruption.
Accessibility signals: Emergency reachability Rj measures whether a facility is reachable within time budget τ, while Tj is shortest-path travel time under hazard-conditioned weights.Evacuation redundancy Kj counts edge-disjoint or node-disjoint feasible routes to the nearest shelter.
Splits and leakage controls: The evaluation uses temporal, spatial-block, and unseen-city splits, including a held-out Baku city and 20% held-out 1 km×1 km grid cells.Leakage controls include record deduplication, future-information exclusion, and training-partition-only normalization.
Evaluation metrics: Metrics cover prediction quality, high-risk recall and false-negative rate, uncertainty calibration, reachability, travel time, and redundancy.Calibration uses reliability diagrams, ECE, and credible-interval coverage; transportation outcomes are compared under baseline and counterfactual scenarios.

6) Edge–Cloud Deployment Model:

Skjold-DiT is designed for deployment in which cloud or edge servers perform multimodal encoding and diffusion sampling, while vehicles consume compact, periodically updated risk and accessibility layers. The BCUR dataset supplies building-level hazard, demographic, infrastructure, and transportation data, with documented harmonization, missingness, provenance, and licensing procedures.

Deployment Model: Cloud or edge servers run heavy multimodal encoding and diffusion sampling, while vehicles query compact risk and accessibility layers for low-latency routing.The layers include per-road-segment risk, per-zone accessibility constraints, hazard-conditioned travel-time weights, and reachability indicators.
Dataset: 847,392 buildings across six cities form BCUR, which includes road networks, public transit accessibility, emergency-service locations, and historical evacuation-route usage.Some municipal and insurance-derived annotations are restricted, with processed extracts and metadata available upon reasonable request.
Annotations: Transportation accessibility is quantified through emergency-vehicle travel times and evacuation-route capacity under varied hazard scenarios.The dataset also annotates flood depth, heat stress, and structural vulnerability using geospatial, thermal, historical-damage, and network-analysis inputs.
Preprocessing: BCUR layers are harmonized to a common building index, coordinate reference system, and yearly forecasting horizon.Raster features are summarized over building footprints, while time-varying channels are aligned to yearly bins.
Architecture: Norrland-Fusion uses specialized imagery, tabular, graph, and time-series encoders before cross-modal attention and concatenation into a shared representation.The architecture supports 10-30% random modality masking during training to improve robustness when transportation data are incomplete.

2) Diffusion Transformer Backbone:

The diffusion-transformer backbone represents clustered buildings as spatio-temporal tokens and conditions denoising on graph structure, prompts, temporal context, and transportation constraints. Fjell-Prompt supports compositional transfer to unseen cities, while Valkyrie-Forecast uses intervention prompts and probabilistic sampling for counterfactual risk analysis.

Diffusion Transformer Backbone: Buildings are partitioned into K coordinate-based spatial clusters, each represented as a token in a spatio-temporal transformer sequence.The sequence is processed by a 24-layer transformer with 16 attention heads and hidden dimension 1024.
Diffusion Process: The forward diffusion process adds Gaussian noise over T = 1000 steps using a linear schedule with βt ∈ [0.0001, 0.02].The reverse process is conditioned on graph structure, prompts, temporal context, and transportation constraints.
Fjell-Prompt: Fjell-Prompt decomposes hazard scenarios and transportation constraints into hierarchical prompt templates for zero-shot inference in unseen cities.New-city prompts use available metadata without city-specific disaster-history training.
Valkyrie-Forecast: Valkyrie-Forecast modifies building features to reflect policy interventions and samples counterfactual risk trajectories under those changes.Examples include bioswales, elevated foundations, redundant evacuation routes, and household relocation.
Uncertainty Quantification: The model generates 100 samples per building and reports mean predictions with 90% credible intervals for intervention uncertainty.This probabilistic output is intended for risk-averse transportation planning and infrastructure investment decisions.

D. Training Procedure

Training proceeds through multi-task pre-training, cross-city fine-tuning, and zero-shot validation, with evaluation spanning predictive performance, calibration, and counterfactual validity. Zero-shot Baku transfer reaches 87.2% flood accuracy, 7.5% below Copenhagen in-distribution performance, with heat MAE of 2.1°C.

Configuration: Training uses AdamW with weight decay 10^-2, learning rate 2×10^-4, cosine decay, gradient clipping at 1.0, and mixed precision across 5 random seeds.Results are reported as mean±std.
Training Procedure: Stage 1 pre-trains on Copenhagen, Stockholm, and Oslo for flood, heat, structural-damage, and transportation-accessibility tasks.Training uses 200 epochs on 8×A100 GPUs with batch size 128.
Training Procedure: Stage 2 fine-tunes on Riga and Tallinn while freezing modality encoders and the first 12 transformer layers.The last 12 transformer layers and task-specific heads remain trainable for 50 epochs per city.
Training Procedure: Stage 3 validates zero-shot on Baku without training data, using city metadata and validation events from the 2010 Kura flood and 2024 heatwave.This tests prompt-based inference in a city excluded from training.

IV. EXPERIMENTAL RESULTS

Skjold-DiT is evaluated across temporal, cross-city, calibration, long-horizon, validation, and counterfactual settings, with results emphasizing prediction quality, transfer, uncertainty reliability, and equity-aware planning support.

Flood Risk Prediction Performance: 94.7% accuracy in 10-year flood-risk classification outperforms specialized flood models by 6.5% absolute.False negative rate is reduced 67% versus the physics-based HAND-DEM approach.
Multi-City Generalization: 87.2% flood accuracy transfers to Baku zero-shot, only 7.5% below Copenhagen in-distribution performance.Heat MAE is 2.1°C without Baku-specific training data.
Multi-City Generalization: 85.3% of flooded structures are correctly identified in validation against the 2010 Kura River flood, versus 68% for insurance risk models.The event documentation covers 14,287 damaged buildings.
Long-Term Forecast Accuracy: 86% accuracy is maintained at the 10-year forecast horizon while uncertainty remains calibrated under the temporal split.Comparison methods degrade beyond 3 years, limiting practical utility for transportation infrastructure planning.
Calibration: 90% credible intervals contain ground truth in 91.2% of test instances, with ECE of 0.037.These probabilistic outputs are evaluated for risk-sensitive routing and evacuation planning.
Counterfactual Policy Impact: The Integrated Plan produces the largest predicted 10-year flood-risk reduction among tested Copenhagen interventions and supports prioritizing neighborhoods where risk and accessibility constraints co-occur.Counterfactual outcomes are intended for comparing intervention portfolios and planning, rather than verified realized impacts.

G. Retrospective Case Study (Copenhagen)

The Copenhagen case study evaluates whether predicted risk hotspots align with incidents and how counterfactual scenarios affect transportation accessibility indicators, while limiting claims about realized operational benefits. Results include component ablations, feature importance, calibration, and Baku transfer findings with emergency-access implications.

Case-study scope: The case study compares predicted risk hotspots with observed incident reports and examines counterfactual changes in emergency reachability under road-impassability constraints.It is presented as a reproducible evaluation and decision-support workflow rather than evidence of realized monetary savings or operational performance changes.
Error analysis: The evaluation stratifies errors by hazard severity, modal-data availability, and transportation-network density, and recommends uncertainty thresholds, accessibility consistency checks, and periodic backtesting.These recommendations address deployment-oriented handling of uncertainty and accessibility-layer behavior.
Ablation findings: 13.4% Baku improvement from prompt engineering validates zero-shot transfer in the ablation study.Cross-modal attention adds 3.5% accuracy, graph encoding adds 4.9%, and transportation integration adds 2.1%.
Feature importance: Elevation/topography contributes 28% of flood-prediction importance, followed by historical climate events at 19% and building structure at 16%.Infrastructure proximity contributes 14%, demographics 12%, satellite imagery 11%, and transportation features 8% with importance for emergency accessibility scoring.
Calibration: ECE: 0.037 indicates that predicted probabilities closely match empirical frequencies across confidence bins.The reliability diagram uses blue prediction bars, a black perfect-calibration diagonal, and red dotted calibration gaps.
Baku transfer case: 14,287 Baku buildings, or 10.8% of city stock, exceed 20% 10-year flood probability and house 47,382 residents with limited emergency vehicle access.Additional findings identify 8,942 heat-vulnerable buildings, 2,184 buildings facing both hazards, and concentrated vulnerability in Sabunchu, Yasamal, and Nizami.

B. Transportation-Aware Policy Recommendations

The recommendations connect predicted climate-risk and accessibility layers to transportation planning, routing, dispatch, and resilience investments. They emphasize coordinated building retrofits, infrastructure standards, scenario screening, reproducible evaluation, and subgroup robustness analysis.

Priority interventions: Priority interventions designate high-risk buildings as resilience zones while co-optimizing retrofit decisions with redundant emergency access routes.Proposed measures include subsidized elevated foundations and waterproofing, resilient redevelopment permits, green-infrastructure incentives, and prioritized access improvements.
Building standards: Recommended standards combine flood-safe foundation elevations, heat-reflective roofs, rainwater harvesting where feasible, and emergency-vehicle access compliance.The standards target both housing resilience and continued vehicle accessibility.
Scenario planning: Scenario screening uses model outputs to quantify accessibility indicators under hazards and identify candidate evacuation-route upgrades.This workflow links neighborhood-scale interventions to transportation-relevant planning questions.
Reproducibility: The reproducibility package includes a data card, evaluation scripts, fixed seeds, and multi-seed reporting with mean and dispersion for key metrics.The data card covers modalities, provenance, licensing, preprocessing, missingness, and label generation.
Operational evaluation: Deployment evaluation reports training compute, sampler inference cost, end-to-end layer-generation latency, and routing-time latency for precomputed hazard-conditioned weights.These measurements cover both model production and vehicle queries of reachability indicators.
Equity and robustness: Subgroup performance, calibration, and cross-city distribution shifts are reported across income quintiles and high- versus low-accessibility neighborhoods.The analysis is designed to characterize robustness and detect disparities across socio-economic and accessibility strata.

D. Misuse, Safety, and Policy Risk

The paper frames Skjold-DiT as a tool for transportation-aware climate-risk prediction and intervention analysis, but highlights misuse risks, modeling limitations, and the need for guarded deployment. Its outputs support accessibility-focused planning while remaining bounded by static, historical, and aggregated assumptions.

Misuse and safety: Building-level risk predictions may enable discriminatory pricing or exclusion, so high-stakes emergency-routing use should include human oversight and prospective validation.The paper also recommends avoiding direct identifiers, aggregating public-facing tools, and reporting uncertainty and calibration.
Spatial granularity: The framework aggregates buildings into 100m spatial clusters, potentially missing intra-block variation.Future work proposes building-level graph neural networks while maintaining scalability.
Infrastructure dynamics: Infrastructure is modeled with static features rather than dynamic stormwater and evacuation-traffic simulations, constraining capacity estimation.Future versions are intended to integrate dynamic network simulation.
Human behavior: Counterfactual simulations assume static occupancy patterns, limiting realism for evacuation, migration, and adaptive behavior.Agent-based behavioral models are proposed to improve emergency-response planning.
Cascading hazards: The current framework treats flood and heat independently, leaving cascading failures such as outages and transportation breakdowns outside the model.Integration with critical-infrastructure simulators is identified as necessary for cascading-effect modeling.
Real-time deployment: The system operates on historical data, so real-time integration with sensors, intelligent vehicles, and weather monitoring remains future work.The paper identifies nowcasting and dynamic network-capacity models as next steps.
Supported scope: Skjold-DiT links multi-hazard building-risk signals with transportation accessibility and probabilistic intervention reasoning for intelligent-transportation workflows.The key outputs include calibrated reachability, travel-time inflation, and route redundancy constraints.