TerrainFloodSense: Improving seamless flood mapping with cloudy satellite imagery via water occurrence and terrain data fusion

• Seamless flood mapping using optical satellite imagery under cloud cover. • Bayesian fusion of water occurrence and terrain data enhances cloud reconstruction. • Up to 8.86 % absolute OA increase in cloud reconstruction for extreme flooding. • Evident benefits of cloud reconstruction in flood monitoring and impact assessment. • Leveraging optical satellite images for flood mapping in cloudy environments. Extreme flood disasters are intensified by climate change, exposing an increasing share of the global population to flood hazards. Accurate monitoring of inundation extents during floods is crucial for disaster management and impact assessment. While remote sensing can provide strong support for flood monitoring, optical satellite images often face significant challenges due to weather conditions and infrequent revisits, particularly in cloudy and rainy regions. To address this limitation and achieve seamless flood mapping with cloudy satellite images, this paper proposes TerrainFloodSense, a novel method that fuses water occurrence with terrain data to enhance the reconstruction of cloud-covered flooding areas, especially under extreme and unprecedented flood scenarios. Specifically, TerrainFloodSense first generates enhanced water occurrence data by Bayesian fusion of terrain indices, including Digital Surface Model (DSM) along with Height Above the Nearest Drainage (HAND), and historical water occurrence data. Then, enhanced water occurrence data are used to fill gaps caused by clouds in water maps derived from optical satellite images, guided by the submaximal stability assumption. The basic idea is that prior terrain information can be incorporated into the initial water occurrence data to enhance the ability to predict the inundation probabilities for both regular pre-flood water and extreme floodwater and to help reconstruction of cloud-covered flooding areas even under extreme flooding scenarios. Simulated experiments and applications in large-area flood mapping cases confirmed that TerrainFloodSense significantly outperformed existing methods, achieving absolute accuracy improvements of 2.95%–8.86% in overall accuracy and 0.038–0.087 increases in F1-Score under extreme flooding scenarios. This study demonstrated that the fusion of water occurrence and terrain data can effectively improve seamless flood mapping by using optical satellite images, supporting disaster monitoring and impact assessment in cloudy and rainy environments. The code associated with this study has been made publicly accessible via https://github.com/RCAIG/TerrainFloodSense .