FURI | Spring 2026

Few-Shot Disaster Classification via Causal Disentanglement and Cross-Modal Flattening

Rapid situational awareness is vital for public safety during emergencies, yet current disaster classification models require massive datasets to function reliably. In the early stages of a unique or localized crisis, labeled data is often non-existent, and traditional deep learning models fail to categorize these events accurately with limited samples. This “data hunger” creates a critical bottleneck, preventing intelligent systems from providing immediate support when every second counts.

To address this, we propose a few-shot multimodal learning framework for disaster classification. Our approach utilizes causal disentanglement to separate invariant disaster features from noise and cross-modal flattening to align information across different data types (such as text and images). By identifying the underlying causal mechanisms of a crisis rather than relying on volume-heavy pattern matching, this framework enables “off-the-shelf” models to adapt to new disaster types with minimal supervision. This research contributes to national resilience by ensuring emergency systems remain accurate and deployable even in data-poor scenarios.