Multimodal Deep Learning for Disaster Classification from Social Media Data in Brazil and Peru

This paper compares the performance of pre-trained and fine-tuned Vision-Language Models (SigLIP2 Naflex and SigLIP Multilingual) for classifying disasters in images scraped from social media in Peru and Brazil. While zero-shot inference showed limited results, fine-tuning both models improved performance, with SigLIP2 slightly outperforming the others. The work highlights areas for further improvement in disaster detection models for Latin America.