Replication data for: Towards Multimodal Geospatial Reasoning: A Foundation Model Approach for Disaster Detection from Social Media, News, and Weather Data

This dataset provides a time-stamped H3 hexagonal grid for benchmarking multimodal disaster detection models across two events: (1) the September 2024 floods in Austria and Czechia, and (2) the January 2025 Southern California wildfires. It integrates satellite-derived disaster footprints from DLR and NASA to indicate the presence of an active flood or wildfire. In addition, it includes aggregated statistics from Bluesky social media posts, GDELT news headlines, and daily weather observations from Meteostat. All related data collection scripts for social media, news, and weather are available on GitHub. This dataset dataset can be used directly or adapted as a benchmark for multimodal disaster detection across diverse data sources. It can also be extended to incorporate or evaluate additional modalities. Data schema Each record represents one H3 cell on a specific date and includes the following columns. H3 grid and metadata event: Disaster event identifier date: Observation date (UTC) h3_index: H3 cell index h3_polygon: Polygon of the H3 cell h3_resolution: H3 resolution level h3_centroid: Cell centroid centroid_location: Human-readable location address event_type: Type of disaster (e.g., flood, wildfire) Satellite-derived disaster presence ground_truth_count: Number of hazard-affected blobs ground_truth_presence: Binary indicator of disaster presence Aggregated Bluesky statistics bsky_posts_count bsky_posts_disaster_related_count ratio_bsky_posts_disaster_related Aggregated GDELT statistics gdelt_articles_count gdelt_articles_disaster_related_count ratio_gdelt_articles_disaster_related Weather data (Meteostat) tavg: Average temperature tmin: Minimum temperature tmax: Maximum temperature prcp: Precipitation wspd: Wind speed