Evaluating the effectiveness of NDWI in SWE-GNN-based flood modelling for the Pemali River Basin, Brebes Regency, Indonesia

Abstract Indonesia, as a tropical country straddling the equator, has two seasons, with high rainfall during the rainy season often triggering flooding in various regions. Brebes Regency is particularly vulnerable to flooding due to the overflowing Pemali River, exacerbated by flat topography, inadequate drainage, and alluvial soil with low infiltration capacity. These conditions emphasize the need to develop accurate and efficient flood prediction models to support risk mitigation. The development of flood modeling technology has shifted from conventional hydrodynamic methods that require intensive computation to the application of machine learning (ML) and deep learning (DL). This research implements the Shallow Water Equation-Graph Neural Network (SWE-GNN), an innovative approach that integrates shallow water equations with the ability of Graph Neural Networks to capture spatial dependencies. SWE-GNN is able to efficiently model flood dynamics at low computational cost and has better generalization than traditional models. The model was developed using the parameters of Digital Elevation Model (DEM), slope gradient, rainfall, soil type, land use (Land Use/Land Cover), and Normalized Difference Water Index (NDWI). This study specifically assesses the contribution of NDWI in improving the accuracy of SWE-GNN-based flood predictions. The resulting findings are expected to not only strengthen artificial intelligence-based flood modeling, but also provide a scientific basis for disaster mitigation strategies, spatial planning, and flood risk reduction in disaster-prone areas.