Quantifying Damage from Extreme Events with AI

We develop models to quantify damage caused by extreme events such as floods, earthquakes, and wildfires using satellite imagery, neural networks, large language models, and vision language models. Our innovative approach enables damage forecasting even in the absence of historical data. This capability provides stakeholders with valuable insights to proactively prepare for potential threats and mitigate risks associated with extreme events.

A Multimodal, Multilingual, and Multidimensional Pipeline for Fine-grained Crowdsourcing Earthquake Damage Evaluation
Z. Ma, L. Li, J. Li, W. Hua, Q. Feng, and Y. Miura*. (2025). Under Review.

We introduce a 3M (Multimodal, Multilingual, Multidimensional) pipeline that uses multimodal large language models (MLLMs) to assess disaster impacts from social media. Evaluated across major earthquake events in two countries, our approach integrates image and text data to provide timely, fine-grained damage assessments that correlate with seismic ground truth, highlighting the potential of MLLMs for real-time crisis response.

RoofNet: A Global Multimodal Dataset for Roof Material Classification
N. Law and Y. Miura. (2025). Under Review. [DOI] [GitHub]

We present RoofNet, the first geographically diverse dataset for global roof material classification. With over 51,500 EO image-text pairs from 184 sites and 112 countries, RoofNet enables scalable, AI-driven risk assessment of building vulnerability to natural hazards. Fine-tuned with a vision-language model (VLM) using expert annotations and prompt tuning, it supports downstream tasks in disaster preparedness, insurance, and infrastructure resilience.

PREDISM: Pre-Disaster Modelling with CNN ensembles for at-risk communities
V. Anand and Y. Miura. NeurIPS 2021 Tackling Climate Change with Machine Learning Workshop (2021). [DOI]

With rising natural hazards, the machine learning community is focusing more on disaster damage but less on preemptive mitigation. This study presents PREDISM, a model using ResNets and decision trees to predict building-level damage before disasters, improving resource allocation to reduce losses. PREDISM allows damage forecast in the areas without requiring post-hazard damage data.