Multimodal Earthquake Damage Evaluation
Rapid earthquake assessment is limited by sparse ground sensors and delayed official reporting. This project investigates whether multimodal, multilingual social-media observations can provide complementary, fine-grained evidence immediately after an event.
Technical innovation
The Multimodal, Multilingual, and Multidimensional (3M) pipeline uses foundation vision-language models to align text, images, spatial context, and damage dimensions. It evaluates three models across the 2019 Ridgecrest and 2021 Fukushima earthquakes against ground-truth seismic data.
Practical insight
The models show strong potential for event localization and image-text fusion, while also revealing sensitivity to language, modality, and epicentral distance. These findings clarify both how multimodal AI can support rapid assessment and where expert validation remains essential.