🌀 3M Pipeline for Disaster Damage Assessment

This repository implements the 3M (Multimodal, Multilingual, and Multidimensional) pipeline for fine-grained disaster damage assessment using social media and multimodal large language models (MLLMs).

🔍 Overview

The 3M pipeline operates in three stages:

1. Data Preparation: Filtering and geolocating disaster-related tweets.
2. Damage Evaluation: Using MLLMs to classify damage severity (MMI scale) from text and image inputs.
3. Model Evaluation: Correlating model predictions with DYFI ground-truth data and analyzing input modality, prompt sensitivity, and reasoning transparency.

🌍 Case Studies

• 2019 Ridgecrest Earthquake (USA)
• 2021 Fukushima Earthquake (Japan)

🤖 Supported Models

• LLaVA 3–8B
• Qwen 2.5-VL-7B
• Gemini-2.5-Flash

📁 Directory Structure

├── data_opreparation/     # Preprocessed tweet data  
├── damage_evaluation/     # Model call scripts and configs
├── prompt/                # Prompt templates used for LLMs  
├── model_validation/      # Correlation and reasoning analysis
├── image/                 # framework and key result visualizations 
├── results/               # CSV files with sample damage evaluation results  
└── README.md              # Project documentation  

📌 Note: All original user information has been removed from these files. The full dataset is available upon request.

📊 Results

Key findings from the 3M pipeline experiments:

• ~Near-moderate correlation with DYFI ground-truth seismic data
• Robust performance in urban and multilingual contexts
• Effective reasoning patterns and model interpretability analysis
• Limitations in high-intensity damage detection and low-signal/multilingual regions