Comparative analysis of open-source SAR-based flood datasets for accurate flood extent delineation using deep learning

Kurzfassung

Climate change has prominently modified the global hydrologic cycle in recent years, resulting to a constant growth in frequency and severity of rainfall in various regions around the world. One of the cascading effects of this phenomenon is the persistent rise in the mean sea level, increasing the likelihood of catastrophic floods in low-lying areas. Extreme weather events, such as typhoons and sustained precipitation, not only exacerbate flooding in coastal zones but also leave immense damage in urban areas. Therefore, precise flood extent detection is crucial for the creation of effective early warning systems, disaster response, and recovery planning. The identification of flooded areas used to be a labor-intensive and time-consuming task that requires human expertise. With the application of Earth observation (EO) data using artificial intelligence, deep learning models were developed to identify flooded areas from EO datasets of different modalities. However, the training process in any deep learning architecture entails a large amount of labeled reference data to be fed into the network for feature extraction. Among the datasets available for flood detection are optical images derived from drone surveys (e.g., FloodNet) or satellite missions (e.g., SpaceNet). Although these datasets have fine resolution suitable for defining flood boundaries, they are usually constrained by their reliance on optical imagery and lack of geographical diversity due to limited spatial coverage. In contrast, some flood datasets utilized synthetic aperture radar (SAR) images due to their ability to penetrate clouds and operate under low light, which are typical conditions during a flood event. An example of which is S1GFloods, a SAR-based dataset featuring flood events in different terrains under various flood-triggering scenarios. This dataset accounted 46 flood events from 2015-2022 resulting to 5,360 Sentinel-1 image pairs. As flood detection in built-up areas imposes greater challenge in the field of remote sensing, a more urban-focused flood dataset was introduced namely the UrbanSARFloods. Similar to S1GFloods, this dataset utilized pre- and post-event Sentinel-1 images, representing 18 flood events across the world. The dataset includes coherence and intensity images derived from Sentinel-1 imagery. Training samples were semi-automatically annotated to demarcate the boundaries between flooded urban and open areas and non-flooded regions. Considering the advantages of using SAR images, extensive geographic coverage, and diversity in flood scenarios, this study aims to evaluate and compare the reliability of SAR-based flood datasets, S1GFloods and UrbanSARFloods, in accurately delineating flood extents in new topographic context. Each dataset was used to independently train a convolutional neural network (CNN) model, keeping hyperparameters uniform across all experiments. The analysis is focused on applying the trained models to a test site in Okayama Prefecture, Japan that was severely devastated by Typhoon Prapiroon in 2018. Pre- and post-event SAR images and their corresponding coherence images of the test site were collected and analyzed, ensuring that the data remained unseen during training. Using ground truth data, performance metrics were determined to evaluate the models’ accuracy in delineating flood extent. Findings highlight the strengths and limitations of using SAR-based flood datasets for flood mapping. While SAR images demonstrates good performance in extracting prominent flood features, the geographical distribution of sample datasets remains crucial in applying the model to a different site. In conclusion, this research emphasizes the importance of integrating multiple data sources for improved flood detection accuracy and establishes a framework for leveraging CNN in flood mapping applications. Future work will include the use of multimodal datasets covering wider geographic span and explore its applicability to other flooding scenarios with diverse topographical characteristics.