Solmaz Khazaei: Flood detection using deep learning-based segmentation

Flood is the most common natural disaster in the world, and can have catastrophic impacts on human society and the environment, including infrastructure damage, agricultural losses, and casualties, resulting in widespread economic and social disruptions. 

In early studies, water body detection relied on on-the-spot investigation, hydrological models and common remote sensing techniques that face issues like slow processing and real-time delays. By addressing this challenges we propose a novel hybrid PoLSAR-metaheuristic-DL models and high-resolution remote sensing data to generate accurate and rapid flood mapping for one of the huge recent flood in France.

Compared with standard synthetic aperture radars (SAR), polarimetric synthetic aperture radar (PolSAR) is an advanced technique of SAR remote sensing. So, by using polarimetric decomposition methods, features were extracted and feature selection problem, one of the most challenging, was solved by using metaheuristic techniques. The selected features fed into three deep learning-based segmentation models- U_Net_V3, Nested_UNet and Efficient_UNet. 

The reliability of the generated flood maps was evaluated using Accuracy, precision and recall metrics. Our experimental results indicate that Nested_UNet integrate with optimized PolSAR data achieves the highest segmentation performance, with an accuracy of 0.910, precision of 0.914, and recall of 0.909. 

These findings underscore the capability of Nested_UNet, demonstrates superior feature extraction abilities, making it a promising choice for real-time flood segmentation applications. Moreover, detecting the knowledge of flooded areas, officials can actively adopt steps to reduce the potential impact of flood, ensure the sustainable management of natural resources and mitigate flood impacts.

Solmaz is a PhD student at KTH Royal Institute of Technology, working at the intersection of machine learning and Earth observation. Her research focuses on developing deep learning methods for environmental monitoring and disaster analysis, particularly using satellite data such as SAR imagery for flood detection. She is particularly interested in transfer learning and building models that generalize across different geographic regions. Her work aims to support climate resilience through scalable AI-driven geospatial solutions.