FlowForm: AI Framework for Satellite Flood Synthesis

A new framework named FlowForm has been developed by researchers to synthesize satellite images of floods, tackling the challenge of limited high-quality paired data for flood assessment models. This innovative approach combines latent regularization inspired by shallow water equations with structure-aware conditioning to produce authentic flood visuals. FlowForm features two main elements: the Flood Descriptor Module (FDM), which applies differentiable penalties on the residuals of the steady-state Shallow Water Equation within auxiliary latent fields at the diffusion bottleneck, and the Terrain Anchor Adapter (TAA), which incorporates depth, semantic, and edge features across four encoder scales of the U-Net. The team also created FloodScape, an extensive high-resolution dataset of paired satellite images captured before and after disasters. Detailed in a paper on arXiv (arXiv:2608.03822), this cross-type submission aims to enhance the realism of spatial layouts and maintain scene structures in generated flood imagery, addressing distortions found in current methods. This advancement is crucial for improving flood assessment models through enhanced data augmentation.