Topology-Informed Neural Networks Boost Flood Detection in Satellite Imagery

Floods cause widespread devastation, but rapid detection from space is often hindered by cloud cover that blocks optical satellite sensors. A new paper from researchers Sophia Li, Max Zhao, Raghu G. Raj, and Tianyu Chen tackles this by introducing topology-informed neural networks. The work, published on arXiv, leverages topological data analysis (TDA) — a mathematical framework that captures global structural features — to enhance flood detection in both optical and synthetic aperture radar (SAR) imagery.

Existing deep learning approaches, such as ResNet-50 combined with GRU networks or vision transformers, have shown strong results on the SEN12-FLOOD benchmark dataset but often act as opaque black boxes. The authors systematically evaluate topological descriptors as complementary features, demonstrating that TDA carries meaningful flood signals independently and can boost accuracy and interpretability when integrated into standard neural network architectures. This is especially valuable for safety-critical applications like disaster response.

By adding topological features, the system becomes more robust to cloudy conditions that plague optical sensors, while remaining interpretable for analysts. The open-source SEN12-FLOOD dataset, which includes both optical and SAR imagery, provides a realistic testbed. Although the paper does not report specific performance numbers, it positions TDA as a powerful tool for environmental monitoring, potentially enabling faster and more trustworthy flood mapping in real-world scenarios.