IMAFD: An Interpretable Multi-stage Approach to Flood Detection from time series Multispectral Data

Multi-stage approach improves efficiency: IMAFD reduces computational costs by filtering satellite images through four sequential stages before performing detailed analysis, making large-scale flood detection more practical and resource-efficient compared to processing all images equally.

Interpretability built into the system, not added afterward: Unlike most AI explainability methods that explain decisions after the fact, this approach embeds interpretable decision-making directly into the flood detection process through semantic segmentation, allowing users to understand why floods are detected at each stage.

Combines multiple detection methods for accuracy: The system uses both anomaly detection (identifying unusual patterns in image sequences and individual pixels) and semantic segmentation (classifying areas as land, water, or cloud) to cross-verify flood detection, reducing false alarms.

Addresses a real-world problem in disaster response: By automating flood detection from satellite imagery while maintaining transparency, this technology could help emergency responders make faster, more informed decisions during flooding events when time is critical.

Demonstrates responsible AI in practice: The paper exemplifies how AI systems handling high-stakes decisions (like disaster detection) can be designed with both efficiency and explainability as core features rather than afterthoughts, supporting human decision-makers rather than replacing their judgment.