SARFlood: A Web-Based, Cloud-Native Platform for Automated and Optimized ML-based SAR Flood Mapping

Synthetic Aperture Radar (SAR) imagery offers weather-independent observation capabilities critical for monitoring flood events. However, SAR-based flood detection workflows typically require specialized software, local computational resources, and expert knowledge in remote sensing. This work presents SARFlood, a web-accessible application that automates the complete SAR flood detection pipeline using the OpenEO platform. SARFlood is built on a Flask backend architecture designed for accessibility and reproducibility. Users interact with the system through a web interface that guides them through case study creation, including Area of Interest (AOI) definition via shapefile upload, event date specification, and optional ground truth data integration. The application implements OpenEO OAuth 2.0 authentication using the device code flow, enabling secure access to the Copernicus Data Space Ecosystem (CDSE) backend without requiring users to manage API credentials locally. Session-based project management allows users to track processing progress in real-time through a status reporting system that monitors each pipeline stage. Data acquisition is performed server-side via OpenEO, while feature engineering processors execute locally. The data acquisition module fetches multiple data sources through a unified OpenEO interface: pre-event and post-event Sentinel-1 VV and VH imagery, Digital Elevation Models (DEM) with automatic source fallback (FABDEM, Copernicus 30m/90m), and ESA WorldCover land cover classification. The OpenStreetMap water body features and the FathomDEM are acquired via their own APIs/websites. A caching system prevents redundant API calls for previously acquired datasets, significantly reducing processing time for iterative analyses, while keeping licensing in mind so only users who are logged in and have the according license will be able to access the cached files. The processing pipeline computes a comprehensive feature stack for flood detection. SAR derivatives include intensity bands, VV/VH polarization ratios, and change detection metrics computed in decibel space to enhance flood signal discrimination. Topographic features encompass slope and Height Above Nearest Drainage (HAND) derived from the DEM, as key indicators of flood susceptibility. Flow direction calculations use an expanded bounding box to determine the extended HAND computation domain to address edge artifacts, finally cropped to the original AOI during band compilation, ensuring computationally efficient and accurate flow routing. Additionally, stream burning is implemented to improve drainage network delineation. Further, contextual features include Euclidean Distance to Water and rasterized land cover classification. Users can currently upload ground truth shapefiles (e.g., Copernicus EMS), which are automatically rasterized and compiled into the output stack, enabling supervised classification workflows.  

SARFlood includes integrated sampling and training modules. Multiple strategies such as Simple Random, Stratified, Generalized Random Tessellation Stratified, and Systematic Grid sampling are supported. The training module implements Random Forest classification with Leave-One-Group-Out Cross-Validation across multiple case studies, hyperparameter optimization via Bayesian search, and feature importance assessment through Mean Decrease Impurity, permutation importance, and SHAP values. The platform-, data- and model-agnostic design principles used in developing SARFlood, support open science and FAIR practices in the geoscience community. By combining web accessibility with robust feature engineering and machine learning integration, SARFlood provides researchers with a reproducible platform for generating uncertainty-aware flood labels lowering barriers to use.