PWVGAT: Graph Attention Network for Spatial Inference of GNSS Precipitable Water Vapor Over Irregular Station Networks

Abstract

Precipitable water vapor (PWV) derived from Global Navigation Satellite System (GNSS) stations provides millimeter-level accuracy under all weather conditions but is limited to discrete site locations. Extending these observations into spatially continuous fields remains challenging because conventional interpolation assumes spatial isotropy, pointwise machine learning ignores interstation topology, and convolutional approaches require regular grids incompatible with irregular station networks. This study introduces PWVGAT, an edge-enhanced graph attention network that represents the GNSS station network as a graph and learns anisotropic spatial attention weights through masked self-supervised training. A variable masking strategy enables prediction at unseen locations within the constructed station network. Experiments over 193 quality-controlled stations in California demonstrate that PWVGAT achieves a validation-set RMSE of 1.40 mm ($R^{2} = 0.974$), reducing RMSE by 64.4% relative to inverse distance weighting, 32.7% relative to XGBoost, and 31.7% relative to a standard graph convolutional network across six comparison methods. Independent validation against six radiosonde stations (3933 collocated samples) yields RMSE of 2.12 mm and correlation of 0.979. The reconstructed 0.25° 2-D PWV field is consistent with the large-scale spatial patterns represented by ERA5 total column water vapor (correlation 0.970, RMSE 2.89 mm). These results establish graph-structured spatial modeling as an effective framework for high-resolution PWV mapping from sparse, irregularly distributed GNSS networks.