Artificial Intelligence for flood analysis: first results from the AI4Flood project

Floods are one of the most frequent and the costliest natural disasters. Accurate and rapid mapping of flooded areas becomes more crucial when floods strike densely populated cities. A cost-effective and widely used tool for near real-time flood monitoring is satellite remote sensing. Water can be easily discriminated from other land covers in optical satellite imagery owing to the spectral behavior in visible and infrared ranges of wavelength. However, the optical sensors’ major limitation is their inability to penetrate clouds, resulting in images with missing information and restricting their further use for flood monitoring. In the last decade, Synthetic Aperture Radar (SAR) has played a significant role in operational services for flood management and has been used by agencies worldwide. SAR is an active imaging technique that overcomes problems of optical sensors and provides day and night cloud-free images. The specular reflection from smooth water surfaces makes the water appear black in SAR images. Although SAR amplitude has been widely used operationally in flood detection and monitoring, it is subjected to overestimation of flooded areas, particularly in the arid and semi-arid regions. This is due to the similarity between the radar backscatter over sand and open water surfaces. Interferometric coherence and polarimetric information can overcome this problem by providing complementary information and further refining flood events. Also, advanced machine learning and deep learning approaches have demonstrated the potential to learn from the current data and improve the classification accuracy as well as reduce response time and model development cost. In this study, we present preliminary results of our research obtained within the AI4Flood project (AI for Near Real Time Satellite-based Flood Response), funded within the framework of the 2019 call of Helmholtz AI Projects. The main goal of the project is to improve existing satellite-based emergency mapping methods based on SAR data by training, testing, and validating novel machine learning algorithms that incorporate information from amplitude and coherence of SAR data together with polarimetric decomposition for the semantic segmentation of water bodies in case of flood situations. As the first case study, we focus on southern Iran and present the results obtained for the January 2020 flood event in the arid region of Sistan and Baluchestan. Here we exploit a year-long time series of amplitude, interferometric coherence, and polarimetric decomposition products, such as Entropy (H), and Alpha (α) derived from multi-temporal Sentinel-1 SAR data. In addition, optical imagery from Sentinel-2 was also considered for visualization and validation purposes. We observed that in some areas, the backscatter variations were not high enough to determine the changes during the flood; however, a clear drop-off in coherence was noticed. Similarly, the information stored in H, and α provides complementary information that can be used to detect flooded areas, which are otherwise not possible just by using SAR amplitude. This information, along with Sentinel-2 optical imagery, will be used to train, validate, and test Convolutional Neural Networks (CNNs) to segment permanent and flood water. The aim of the AI4Flood project is to provide a machine learning framework to better detect the flooded areas by using information derived from freely available optical and SAR imagery