Reproduction Code: A Foundation Model Approach for Disaster Detection from Social Media, News, and Weather Data

This repository contains the reproduction code and materials for the paper:

Hanny, D., Dastidar, K.G., Wieland, M., Granitzer, M. & Resch, B. (2026). Towards Multimodal Geospatial Reasoning: A Foundation Model Approach for Disaster Detection from Social Media, News, and Weather Data. [Accepted for publication in Natural Hazards]

📄 Overview

This research introduces a grid-based framework that quantifies disaster detection accuracy relative to satellite-derived reference data. We employ generative Language Models (LMs) to interpret heterogeneous information from Bluesky social media posts, GDELT news headlines, and weather observations through structured prompts and relevance-based data retrieval. The method frames detection as a binary classification problem on an H3 grid.

Our analysis pipeline includes:

1. Data collection: Custom keyword-based crawling of Bluesky posts, GDELT news, and weather observations
2. Data aggregation: Structured aggregation of multimodal data to H3 grid cells
3. Methodology
   1. Statistical anomaly detection: Statistical hotspot and anomaly detection as baseline methods
   2. Foundation Model inference: LM-based interpretation of heterogeneous information sources
4. Evaluation: Systematic comparison against satellite-derived reference
5. Case studies: 2024 Central Europe floods and 2025 Southern California wildfires

📁 Repository Structure

The analysis pipeline is spread across several scripts and Jupyter/Marimo notebooks. A full overview is available below.

├── data/                          # Data files
│   ├── raw/                       # Raw Bluesky, GDELT, and weather data
│   │   ├── 2020_california_wildfires/
│   │   ├── 2024_central_europe_floods/
│   │   ├── 2025_socal_wildfires/
│   │   ├── auxiliary/              # Additional reference data
│   │   └── dlr/                    # Satellite reference data
│   ├── processed/                 # Processed datasets
│   ├── results/                   # Evaluation results
│   └── mapping_data/              # Geospatial reference data
│
├── notebooks/                     # Analysis workflow (run in ascending order)
│   ├── 01_bsky_data_collection.ipynb    # Bluesky data collection
│   ├── 02_esda/                         # Exploratory spatial data analysis
│   ├── 02_ground_truth_prep.py          # Ground truth preparation
│   ├── 03_data_aggregation.py          # Data aggregation to H3 grid
│   ├── 04_statistical_baseline.py      # Statistical baseline methods
│   ├── 05_prompt_optimisation.py        # Prompt optimization experiments
│   ├── 06_few_shot_selection.py         # Few-shot example selection
│   ├── 07_in_context_learning.py        # Main LLM inference pipeline
│   ├── 08_ablation_study.py             # Ablation experiments
│   ├── 09_mixed_ensemble.py             # Ensemble methods
│   ├── 10_visualisation.py             # Result visualization
│   ├── 11_data_anonymisation.py         # Data anonymization
│   └── 13_rev_*.py                     # Additional experiments during paper revisions
│
├── scripts/                       # Non-interactive scripts
│   ├── crawling/                  # Bluesky and GDELT data collection scripts
│   ├── geoparsing/                # Location extraction from text
│   ├── get_weather_data.py        # Weather data retrieval
│   └── in_context_inference.py    # LLM inference helper
│
├── src/                           # Helper modules and reusable functions
│   ├── bsky_search.py             # Bluesky crawling algorithm
│   ├── ensemble.py                # Model ensemble methods
│   ├── eval_metrics.py            # H3 grid-based evaluation metrics
│   ├── helpers.py                 # Data processing utilities
│   ├── validation.py              # Validation functions
│   ├── visualisation.py           # Visualization functions
│   ├── hotspot/                   # Hotspot detection baselines
│   ├── in_context_learning/       # LLM prompt templates
│   ├── irchel_geoparser/          # Geoparsing tools
│   └── nlp/                       # NLP processing utilities
│
├── prompts/                       # LLM prompt templates
├── figures/                       # Generated visualizations
├── logs/                          # Log files
├── Dockerfile
├── docker-compose.yml
├── requirements.txt
└── README.md

⚙️ Getting Started

To reproduce the experiments, we recommend using Docker for a consistent environment. The individual notebooks can be run using marimo as follows:

docker compose run --rm --service-ports marimo

This will start a Marimo notebook server at localhost:8080. Alternatively, you can run the notebooks directly as Python scripts, though the marimo interface is recommended.

For LM-based inference, a running Ollama instance on localhost:11434 or an OpenAI key stored as OPENAI_API_KEY environment variable are furthermore required. Please pull all desired models before running the script.

📊 Data Availability

The primary datasets supporting the conclusions of this article are available in the repository on Zenodo under the DOI 10.5281/zenodo.20038116.

📖 Citation

If you use this code or material in your research, please cite our work accordingly.

@article{Hanny.2026,
  title     = {Towards Multimodal Geospatial Reasoning: A Foundation Model Approach for Disaster Detection from Social Media, News, and Weather Data},
  author    = {Hanny, David and Dastidar, Kanishka Ghosh and Wieland, Marc and Granitzer, Michael and Resch, Bernd},
  journal   = {Natural Hazards},
  year      = {2026}
}