FloodDeepSeek: Topic Classification and Real-Time Monitoring of Flood Events from Social Media Text Data

Flood disasters pose major challenges to emergency management due to their sudden onset, rapid evolution, and strong societal impacts. Conventional monitoring approaches based on physical sensors and remote sensing provide valuable environmental observations but often suffer from latency issues and fail to capture real-time public needs during flood emergencies. To address this gap, this study proposes FloodDeepSeek, a real-time flood monitoring and decision-support framework that leverages social media as a dynamic social sensing source to complement traditional monitoring systems. The framework integrates a distilled large language model (DeepSeek-R1-Distill-Llama-8B) into an event-driven, asynchronous processing pipeline for rapid data ingestion, semantic topic classification, and spatiotemporal visualization of flood-related information. By adopting a queue-based architecture, FloodDeepSeek supports high-throughput, low-latency processing of large-scale social media streams, enabling timely identification and localization of critical messages, such as help-seeking and rescue-related information. A parameter-efficient LoRA fine-tuning strategy is employed to enhance domain adaptation under limited labeled data conditions. Experimental results indicate that the fine-tuned model outperforms a BERT baseline by approximately 2.5% in F1-score for flood-related topic classification. A case study of the July 20, 2021 extreme rainfall event in Henan Province, China demonstrates strong temporal and spatial correspondence between rainfall dynamics and social media activity, highlighting the potential of FloodDeepSeek to support socially aware disaster risk reduction and emergency response.