Method for Meteorological Radiosonde Data Completion Based on Hierarchy-Aware Spatiotemporal Graph Neural Network

Meteorological radiosonde data is an important foundation for weather forecasting and climate research, but it often has missing values due to equipment failures, communication interruptions, etc. The Hierarchy-Aware Spatiotemporal Graph Neural Network (VHT-GNN) introduced in this article achieves high-quality missing value completion and significantly improves accuracy by constructing a three-dimensional graph structure (vertical, horizontal, temporal), combined with hierarchy-aware normalization and edge conditional gating mechanisms.

Radiosonde data covers multiple pressure layers from the ground to an altitude of 30 km, which is crucial for numerical weather prediction, climate monitoring, and extreme weather early warning. However, missing data is widespread. Traditional methods such as Inverse Distance Weighting (IDW) interpolation and linear interpolation struggle to capture complex spatiotemporal correlations; deep learning methods like LSTM and CNN ignore the physical characteristics of the atmospheric vertical structure, leading to limited completion effects.

VHT-GNN represents radiosonde observations as a multi-relational graph, defining three types of edges: vertical edges connecting adjacent pressure layers in the same profile (utilizing physical constraints like hydrostatic equilibrium), horizontal edges connecting different stations at the same pressure layer (utilizing spatial continuity), and temporal edges connecting consecutive time steps of the same station-pressure layer (utilizing temporal inertia), thus capturing the inherent structural characteristics of atmospheric data.

VHT-GNN workflow: Input projection → Vertical/Horizontal/Temporal gated convolution → Adaptive fusion → Masked temporal attention → Output projection. Key technologies: Hierarchy-aware normalization (independent standardization for each pressure layer), edge conditional gating (dynamically adjusting information transfer intensity), adaptive fusion (data-driven weighting), and masked temporal attention (avoiding interference from missing information).

Validated on the IGRA dataset (14 pressure layers, 6 variables), comparing statistical methods (IDW, linear interpolation), deep learning methods (LSTM, CNN), graph neural networks (GraphSAGE, etc.), and ablation models. The results show that VHT-GNN significantly outperforms baselines in accuracy and physical consistency (e.g., hydrostatic equilibrium). The higher the missing rate, the more obvious the advantage, and the stability is good.

1. Numerical weather prediction: Completing data improves the integrity of the initial field and enhances forecasting skills; 2. Climate change reconstruction: Repairing missing historical data to support trend analysis; 3. Extreme weather early warning: Completing interrupted observations under severe weather to provide decision support for forecasters.

The project is divided into data collection (format conversion, merging, analysis) and model implementation (data loading, model definition, training, physical validation) modules. Run examples: cd vht_stgnn; python M99_MAIN.py --model vht_gnn --seed 42 123 456 789 2024; python M98_CompareResults.py to summarize and compare.

Core contributions of VHT-GNN: Three-dimensional graph modeling framework, hierarchy-aware normalization and edge conditional gating, physical consistency check. Future directions: Extending to satellite data completion, joint modeling of multi-source observations, end-to-end physical constraint learning.