Empowering multi-source SAR Flood mapping with unsupervised learning

Abstract Flood mapping plays a crucial role in effective disaster management, risk assessment, and mitigation planning, given the widespread and destructive nature of floods. However, current synthetic aperture radar (SAR)-based methods face challenges related to extensive labeled training data, compromised classification accuracy, and limited applicability across different satellite systems and resolutions. In response to these challenges, our research introduces a pioneering unsupervised SAR-based flood mapping algorithm, inspired by artificial general intelligence principles. Notably, the innovative method demonstrates flexibility, performing effectively across various SAR satellites with differing resolutions and sensors, eliminating the need for satellite-specific algorithms. Our algorithm enhances processing speed and scalability by eliminating labor-intensive labeling of training data and manual intervention. To validate its performance, we conducted tests in three distinct regions using meter-level imagery from HISEA-1, Gaofen-3, and Sentinel-1 satellites. Consistently outperformed prevalent unsupervised techniques like Kmeans and Otsu, and even a Supervised-convolutional neural network segmentation algorithm by AI-Earth, with F1-scores exceeding 0.91. This outstanding performance showcases its accuracy, irrespective of the satellite systems or regions utilized. Furthermore, the seamless integration of our algorithm with high-performance cloud computing platforms such as Google Earth Engine enhances its adaptability and scalability, enabling continuous monitoring of global floods. This is crucial in understanding flood trends, assessing their impacts, and formulating effective disaster mitigation strategies.