Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable Dual-Stream Learning for Local Wind Hazard Prediction in Vulnerable Communities

Mahmuda Akhter Nishu, Chenyu Huang, Milad Roohi, Xin Zhong

TL;DR

This work tackles the lack of localized wind hazard predictions for vulnerable communities by introducing an interpretable dual-stream framework that fuses structured meteorological data with unstructured event narratives. The model combines a Random Forest for numerical features and RoBERTa for narrative text through late fusion, with a lightweight meta-classifier to produce block-level risk outcomes, trained under cross-entropy loss. Key contributions include the novel multimodal architecture, modular training with frozen encoders, and gradient- and ablation-based interpretability analyses that highlight wind speed and wind direction as primary drivers of high-risk predictions. The approach demonstrates strong predictive performance on real-world data from KSUX, achieving high accuracy and robust high-risk detection, thereby offering actionable insights for emergency management and community resilience planning.

Abstract

Wind hazards such as tornadoes and straight-line winds frequently affect vulnerable communities in the Great Plains of the United States, where limited infrastructure and sparse data coverage hinder effective emergency response. Existing forecasting systems focus primarily on meteorological elements and often fail to capture community-specific vulnerabilities, limiting their utility for localized risk assessment and resilience planning. To address this gap, we propose an interpretable dual-stream learning framework that integrates structured numerical weather data with unstructured textual event narratives. Our architecture combines a Random Forest and RoBERTa-based transformer through a late fusion mechanism, enabling robust and context-aware wind hazard prediction. The system is tailored for underserved tribal communities and supports block-level risk assessment. Experimental results show significant performance gains over traditional baselines. Furthermore, gradient-based sensitivity and ablation studies provide insight into the model's decision-making process, enhancing transparency and operational trust. The findings demonstrate both predictive effectiveness and practical value in supporting emergency preparedness and advancing community resilience.

Interpretable Dual-Stream Learning for Local Wind Hazard Prediction in Vulnerable Communities

TL;DR

This work tackles the lack of localized wind hazard predictions for vulnerable communities by introducing an interpretable dual-stream framework that fuses structured meteorological data with unstructured event narratives. The model combines a Random Forest for numerical features and RoBERTa for narrative text through late fusion, with a lightweight meta-classifier to produce block-level risk outcomes, trained under cross-entropy loss. Key contributions include the novel multimodal architecture, modular training with frozen encoders, and gradient- and ablation-based interpretability analyses that highlight wind speed and wind direction as primary drivers of high-risk predictions. The approach demonstrates strong predictive performance on real-world data from KSUX, achieving high accuracy and robust high-risk detection, thereby offering actionable insights for emergency management and community resilience planning.

Abstract

Wind hazards such as tornadoes and straight-line winds frequently affect vulnerable communities in the Great Plains of the United States, where limited infrastructure and sparse data coverage hinder effective emergency response. Existing forecasting systems focus primarily on meteorological elements and often fail to capture community-specific vulnerabilities, limiting their utility for localized risk assessment and resilience planning. To address this gap, we propose an interpretable dual-stream learning framework that integrates structured numerical weather data with unstructured textual event narratives. Our architecture combines a Random Forest and RoBERTa-based transformer through a late fusion mechanism, enabling robust and context-aware wind hazard prediction. The system is tailored for underserved tribal communities and supports block-level risk assessment. Experimental results show significant performance gains over traditional baselines. Furthermore, gradient-based sensitivity and ablation studies provide insight into the model's decision-making process, enhancing transparency and operational trust. The findings demonstrate both predictive effectiveness and practical value in supporting emergency preparedness and advancing community resilience.

Paper Structure

This paper contains 14 sections, 4 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Community Resilience Development of Underserved Communities
  4. AI-Driven Models for Wind Hazard Prediction
  5. Proposed Method
  6. Architecture Overview
  7. Design Motivation and Advantages
  8. Experimental Results
  9. Dataset and Preprocessing
  10. Training and Performance
  11. Comparison
  12. Attribute Importance Analysis
  13. Interpretability in Wind Hazard Prediction
  14. Conclusion

Figures (4)

  • Figure 1: Overview of the proposed dual-stream architecture. Numerical features are processed by a Random Forest, textual narratives by a RoBERTa encoder, and outputs are fused for localized wind hazard prediction.
  • Figure 2: Confusion matrix showing model prediction results for low-risk and high-risk events.
  • Figure 3: Training and validation accuracy curves over epochs.
  • Figure 4: Training and validation loss curves over epochs, indicating stable convergence.