A Sentinel-1/2-based seasonal reference water product to support global flood monitoring activities

Kurzfassung

Floods are in most cases natural processes by which a river overtops its channel and inundates surrounding land thereby sustaining habitats, livelihoods and ecosystem services. In many regions of the world such as in wetlands, coastal and inland delta systems floods are important with respect to food production and biodiversity. However, anomalous inundations can be devastating, causing fatalities and high socio-economic losses. The inundation extent derived from remote sensing data is a key parameter in successful flood disaster management during all phases of the disaster cycle. This information can be derived with increasing frequency and quality due to a steadily growing number of satellite missions in orbit and advances in image analysis. In order to accurately distinguish floods from “normal” hydrologic conditions, up-to-date, high-resolution information on the seasonal water extent is crucial. This information is usually neglected in rapid mapping activities, which may result in a non-reliable representation of the inundation extent, mainly in hydrologically dynamic regions with highly variable intra- and inter-annual availability of surface water. In this work, we provide insights into an automated approach for computing a globally applicable high-resolution reference water product, specifically designed for the use in flood mapping and monitoring applications. The proposed methodology combines deep learning-based modular processing chains developed by the German Aerospace Center (DLR) for water and flood segmentation based on Copernicus Sentinel-1 C-Band Synthetic Aperture Radar (SAR) and Sentinel-2 optical data and calculates permanent as well as monthly seasonal reference water masks over reference time periods of two years. Sentinel-1 Ground Range Detected (GRD) and Sentinel-2 L1C satellite images are automatically analyzed using sensor-specific pre-trained water segmentation models based on the U-Net architecture with EfficientNet encoders. The models have been optimized to deal with the variability of different types and characteristics of water bodies and backgrounds. They have been trained on a globally sampled reference dataset with more than 150,000 256 x 256 pixels tiles. Each sample of the dataset contains triplets of Sentinel-1 and Sentinel-2 images with corresponding quality checked annotations. As during clear-sky conditions a more accurate mapping of water bodies is usually feasible with Sentinel-2 data, this sensor is used as primary source of information for the generation of the reference water product. In regions that are continuously cloud-covered, complementary information from Sentinel-1 is integrated into the classification process. In order to provide information about the quality of the generated reference water masks, we calculate a quality layer, which gives information on the pixel-wise number of valid Sentinel-2 observations covering the derived permanent and seasonal reference water bodies within the selected reference time period. Additionally, the quality layer indicates if a pixel is filled with Sentinel-1 based information in the case that the number of Sentinel-2 observation falls below a critical threshold. The reference water product is demonstrated in several study areas distributed across different climate zones and is systematically cross-compared with already existing external reference water products. Further, the proposed product is applied to real flood events. The results show that the proposed multi-sensor fusion approach is able to generate a consistent reference water product that is suitable for the application in flood disaster response and is capable to produce reasonable results, even if only few or no information from optical data is available. Moreover, the study shows that the consideration of seasonality of water bodies, especially in regions with a strong dynamic of hydrological and climatic conditions, is of high importance as it reduces potential over-estimations of the actual flood inundation extent and gives users a more reliable picture on flood-affected areas. As the product is independent from flood detection algorithms, the proposed reference water product can easily be incorporated in different flood mapping frameworks.