Multimodal retrieval of flood-related remote sensing big data with semantic augmentation from flood events

With the surging adoption of deep learning models, multi-source remote sensing imagery has been increasingly explored to offer a data-driven approach for large-scale flood event analysis and mapping. However, DL-based methods usually suffer from a scarcity of high-quality labeled datasets used for effective model training. Accurately capturing imagery of actual flood-affected areas from massive, dispersed satellite data archives remains significantly challenging under traditional spatiotemporal retrieval paradigms. This challenge arises from insufficient knowledge of flood event semantics, which leads to imprecise spatiotemporal retrieval ranges and difficulty in discriminating flooded areas from permanent water bodies. To address these issues, we propose a semantic-augmented multimodal remote sensing big data retrieval method for analyzing large-scale flood events. It employs a two-phase retrieval strategy: semantic-augmented metadata-based pre-screening across distributed data archives, which refines flood event semantics from the Global Flood Database and Internet news, and DL-enabled multimodal image retrieval for further identification of flood-related imagery, jointly leveraging flood-event-augmented textual semantics and visual features from optical and SAR imagery. Experiments on MFED show that the method achieves an F1-score of 0.965; semantic augmentation improves retrieval accuracy by 11.5%, and multimodal retrieval outperforms single-modality approaches by 28.2%. Across 15 global regions, it reaches 91.2% Top-5 accuracy with a second-level latency.