Graph Neural Network -based identification of homogeneous rainfall regions over Kerala, India

Identifying homogeneous rainfall regions is a fundamental step in regional hydrological analysis. Traditional regionalization approaches often rely on predefined physiographic boundaries or purely statistical clustering, which may inadequately capture complex spatial dependencies in hydroclimatic variables. In regions such as Kerala, India, characterized by complex topography, strong monsoon gradients, and frequent flood events, conventional regionalization methods fail to adequately capture spatial dependence in rainfall variability. This study proposes a Graph Neural Network- based framework for delineating homogeneous rainfall regions to support regional flood frequency analysis and flood risk studies.

Daily gridded rainfall data from the India Meteorological Department (IMD) over Kerala were represented as nodes in a graph, with edges defined by geographical proximity. A two-layer Graph Convolutional Network was trained to learn local rainfall similarity and spatial connectivity. The resulting node embeddings were clustered using the K-means algorithm to identify homogeneous rainfall regions.

Despite using only rainfall information and spatial adjacency, the derived zones closely align with elevation gradients, effectively separating coastal, midland, and western ghats regimes and capturing sharp orographic transitions. This demonstrates that GNN node embeddings can implicitly learn physically meaningful rainfall-topography relationships, providing a robust basis for rainfall regionalization and flood-related hydrocimatic assessments.