Title: Enhancing Flood Inundation Mapping Through Deep Learning-Based Remote Sensing

Abstract:

Floods represent the most widespread natural disaster globally. Accurate and timely mapping of flood inundation is crucial for effective disaster risk management. While optical satellite missions deliver high-resolution, multispectral data essential for detecting floods and mapping inundation, their practical application is significantly hampered by cloud cover during heavy precipitation. Traditional methods for removing clouds, such as temporal compositing or interpolation, frequently struggle to accurately depict the dynamics of flooding.

To address this challenge, this research presents a novel cloud-removal framework for flood imagery utilizing Denoising Diffusion Probabilistic Models, specifically built on a Masked Diffusion Transformer architecture. This approach utilizes self-attention mechanisms to analyze broader spatial contexts and applies masked token modeling to explicitly reconstruct areas obscured by clouds. The model was trained using multispectral Sentinel-2B flood scenes that included realistic cloud patterns, resulting in the generation of cloud-free images that maintain both visual accuracy and hydrological consistency.

The reconstruction capabilities of the model were assessed through standard image quality metrics as well as hydrological measures specific to flood analysis. The findings highlight enhanced continuity of water bodies and the retention of spectral signatures vital for water detection indices. These results suggest that diffusion-based generative modeling serves as a robust and physically consistent solution for cloud removal in optical flood monitoring. This technology facilitates more reliable and continuous observations, thereby supporting better decision-making and disaster risk management in flood-prone areas.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC