Rainfall Spatial Interpolation with Graph Neural Networks

Abstract

Rainfall spatial interpolation is a crucial task to infer rainfall distribution in space for hydrological studies and natural disaster prevention. However, obtaining accurate interpolation results is a non-trivial task due to the complex and dynamic changing spatial correlations of rainfall. Besides, the practical interpolation will be more intractable when there is a lack of auxiliary variables that can help characterize spatial correlations. The performance of traditional interpolation methods is limited by deterministic formulations and statistical assumptions on modeling spatial correlations. Given the huge success of Graph Neural Networks (GNNs), researchers have exploited GNNs for spatial interpolation tasks. However, existing works usually assume the existence of node attributes and rely on a fixed adjacency matrix to guide the message passing among nodes, thus failing to handle practical rainfall interpolation well. To address these limitations, we propose a novel GSI (Graph for Spatial Interpolation) model, which focuses on learning the spatial message-passing mechanism. By constraining the message passing flow and adaptive graph structure learning, GSI can perform effective interpolation by modeling spatial correlations of rainfall adaptively. Extensive experiments show that our approach outperforms the state-of-the-art methods on two real-world raingauge datasets.