Evaluation of open-source SAR-based flood datasets for flood extent mapping in emergency settings

The increasing frequency and severity of flood events, exacerbated by climate change, continue to threaten vulnerable communities worldwide, particularly in low-lying and coastal regions. Timely and accurate flood extent mapping is critical for supporting decision-making and resource allocation during disaster response operations. Advances in machine learning (ML) models, such as deep learning (DL), and the growing availability of Earth observation (EO) data have made automated inundation mapping more viable. However, the reliability of these models is highly dependent on the quality, diversity, and representativeness of the training datasets. Therefore, this study investigates several remote sensing flood datasets, both optical and synthetic aperture radar (SAR), to assess their applicability in emergency response settings. Based on the initial assessment and advantages of using SAR imagery in flood detection, the study evaluates five publicly available Sentinel-1 flood datasets using a DL segmentation model to determine their suitability for emergency response scenarios. Datasets with multi-temporal inputs and high-quality labels achieved superior F1 scores ranging from 0.844 to 0.959. To assess their ability to support emergency operations, the trained models were further tested on three independent flood disasters. While the datasets performed well on their test samples, their predictive accuracy declined substantially when applied to these new flood events, highlighting limitations in current dataset generalizability. The results emphasize that while segmentation models can achieve high accuracy in familiar sites, real-world applications require datasets that capture the heterogeneity of flood scenarios across different regions and conditions. An improved flood dataset design, incorporating multi-temporal and coherence data, a standardized preprocessing framework, and a broader flood geographic distribution, is significant in building reproducible and robust flood detection models. • SAR-based flood datasets were evaluated using a deep learning (DL) segmentation model. • Multi-temporal SAR and coherence data improved flood detection accuracy. • Dataset generalizability is limited by acquisition timing, preprocessing, and geographic coverage. • DL models trained on SAR data enable preliminary flood assessment in real-world emergency scenarios.