FlowForm: Synergizing Fluid Physics with Topological Consistency for Satellite Flood Synthesis
FlowForm is a framework for satellite flood synthesis that integrates SWE-inspired latent regularization with structure-aware conditioning. It includes a Flood Descriptor Module (FDM) that imposes differentiable penalties on residuals of the steady-state Shallow Water Equation in auxiliary latent fields at the diffusion bottleneck, and a Terrain Anchor Adapter (TAA) that injects depth, semantic, and edge features at four encoder scales of the U-Net. The authors curated FloodScape, a large-scale, high-resolution dataset of paired satellite images acquired before and after disasters. FlowForm achieves higher visual fidelity and is evaluated on consistency of flooded regions, zero-shot generalization to a geographically held-out flood event, and sensitivity to individual components.
FlowForm is a novel framework for generating synthetic satellite imagery of floods, addressing the scarcity of high-quality paired data for flood assessment models. It combines fluid physics constraints via the Shallow Water Equation with topological consistency through structure-aware conditioning. The Flood Descriptor Module enforces physical plausibility in latent representations, while the Terrain Anchor Adapter preserves scene structure by incorporating depth, semantic, and edge information. The accompanying FloodScape dataset provides large-scale, high-resolution pre- and post-disaster satellite image pairs. FlowForm demonstrates superior visual fidelity and robust generalization to unseen flood events.
FlowForm's integration of differentiable physics-based penalties (SWE residuals) into the diffusion bottleneck is a key innovation, ensuring that generated flood extents adhere to fluid dynamics. The multi-scale injection of terrain features via TAA likely improves spatial coherence and reduces artifacts. The zero-shot evaluation on a geographically distinct flood event suggests the model learns generalizable flood patterns rather than overfitting to training locations.
This work addresses a critical data bottleneck in disaster response and climate resilience. High-quality synthetic flood imagery can augment training datasets for downstream models used by insurance, government agencies, and humanitarian organizations. The release of FloodScape may become a benchmark for future research in remote sensing and generative AI for environmental monitoring.
FlowForm enables cost-effective generation of realistic flood imagery, reducing reliance on scarce real-world data. This can accelerate the development of flood assessment and prediction models, offering value to insurance risk modeling, urban planning, and emergency response sectors. The FloodScape dataset itself may be licensed or used to train commercial models.
Next signals include potential adoption of FlowForm by remote sensing companies or disaster management agencies for data augmentation. Further research may extend the framework to other natural disasters (e.g., wildfires, landslides) or incorporate temporal dynamics for video generation. Commercialization could involve API-based synthetic data generation services for satellite imagery.