DeepSAR Flood Mapper: Global Flood Mapping on Google Earth Engine Cloud Platform Using MLP Deep Learning Model with Sentinel-1 SAR Imagery and HAND Topographic Data

Appalachian State University Rapid and accurate flood mapping and monitoring are essential for effective disaster management, response, and recovery, as well as for advancing hydrological sciences. Despite the increasing availability of geospatial data and cloud platforms like Google Earth Engine (GEE), existing GEE-based flood mapping applications largely depend on traditional thresholding techniques requiring manual input or struggle with latency due to loose-coupling of complex models. This study introduces DeepSAR Flood Mapper, a novel, fully automated deep learning-based flood mapping application on the Google Earth Engine cloud platform, providing interactive and near-real-time capabilities globally. DeepSAR Flood Mapper utilizes a pretrained Multilayer Perceptron (MLP) deep learning model, selected for its computational efficiency and ability to model highly nonlinear functions, facilitating seamless integration with GEE. The model integrates two crucial input datasets: Sentinel-1 Synthetic Aperture Radar (SAR) imagery (VV and VH polarization) for all-weather surface water detection, and Height Above the Nearest Drainage (HAND) topographic data to mitigate commission errors in elevated areas and enhance reliability. Trained on a combination of global benchmark datasets and historical flood maps, the MLP model is deployed using an Offline Training and Online Prediction (OTOP) coupling strategy, which eliminates data transfer bottlenecks and allows for seamless, on-demand prediction within GEE. The application features an intuitive user interface that allows users to define an Area of Interest and target date, requiring no specialized knowledge. Performance evaluation demonstrates that DeepSAR Flood Mapper significantly improves flood mapping accuracy compared to traditional approaches, including Otsu's thresholding and classical machine learning models, Support Vector Machines (SVM) and Random Forests (RF). Its near-real-time capability supports timely and scalable flood monitoring across diverse geographic regions worldwide.