In this week's Helmholtz AI project showcase, researchers from the GFZ German Research Centre for Geosciences and the German Aerospace Center (DLR) are working on an automated, near-real-time prediction of flood zones with the help of deep learning methods.

Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?

My name is Mahdi Motagh and I am a professor for radar remote sensing at GFZ German Research Centre for Geosciences in Potsdam. My main research is concerned with the use of remote sensing data, in particular observations obtained from Synthetic Aperture Radar (SAR) images to study dynamic changes associated with geophysical and engineering applications such as earthquakes, volcanoes, landslides, floods, city subsidence/uplift, and stability of infrastructures (e.g. dams, bridges, buildings, …). Together with Sandro Martinis from the German Aerospace Center (DLR), I am working on the project 'AI4Flood: AI for Emergency Flood Mapping'.

In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?

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. In the last decade, Synthetic Aperture Radar (SAR) has played a significant role in operational services for flood management and has been used by public agencies worldwide. SAR is an active imaging technique that overcomes problems of optical sensors and provides day and night images regardless of the weather conditions. In particular, the new surge in the availability of freely accessible SAR data via Sentinel-1 satellites has provided golden opportunities to use SAR sensors as an operational instrument for flood monitoring.  The main aim of our project is to develop, using Artificial Intelligence and data-driven approaches, an automated system capable of extracting and detecting flooded areas in near-real-time for generation of flood maps for rapid response activities in case of flood emergencies.

The research will have a major impact for operational flood monitoring and complement the already existing Flood Service at German Aerospace Center (DLR), designed for the automatic near-real-time monitoring of rural flooded surfaces using hierarchical tile-based thresholding and fuzzy logic-based post-classification refinement. Potential risks that may limit the viability of the proposed application include: topography, due to geometrical image distortions related to layover and shadow effects in SAR images, and moderate resolution of Sentinel-1 SAR sensors (10 m) that makes flood detection in urban areas a very challenging task. Another risk that we face is whether a deep learning model trained for certain geographical areas will be able to reliably predict floods in completely different areas and what methodological developments and/or observational constraints need to be done to further improve this performance. 

How important has the Helmholtz AI funding and platform been to carry out this project?

Recent advances in computer vision and the rapid increase of commercially and publicly available medium and high resolution satellite imagery have given rise to a new era of research at the interface between machine learning and remote sensing. The funding from Helmholtz AI was very instrumental in this project as it helped us to explore a variety of global benchmark datasets that can be used for automatic flood monitoring using deep learning and the challenges that we face in this regard. With the experience that we gained and the methodology that we developed, we participated at the ETCI 2021 Competition on Flood Detection organised by the NASA’s Inter Agency Implementation and Advanced Concepts Team (IMPACT). In this competition, we proposed a convolutional neural network (CNN) based on the Unet architecture with a backbone of EfficientNetb7, which was trained with the competition dataset. The performance of our model was then evaluated using several training, testing, and validation methods. In the tests, the model developed by the AI4Flood team achieved an average IOU score of 75.06% and an F-Score of 74.98%. This result placed us among the top three algorithms in the 2021 NASA-ETCI competition (read more here).

Figure: The SAR composite image, the corresponding ground truth and the prediction results for both the UNet and the FPN models for a test dataset.