Graph Neural Network–Driven Spatial Dependency Modeling for Multivariate Environmental Indicator Regression

Abstract

This paper addresses the challenges of insufficient spatial dependency modeling and difficulty in capturing temporal dynamics in multivariate environmental indicator prediction. A spatiotemporal joint modeling method based on graph neural networks is proposed. The method represents the environmental monitoring network as a graph and uses graph convolution to extract non-Euclidean spatial relationships among monitoring sites, capturing topological dependencies between multivariate indicators. A gated recurrent unit is then introduced to model temporal features and characterize the evolution patterns of environmental variables over continuous time windows. The overall architecture integrates graph structure information with temporal dynamics through a structure-aware mechanism, enhancing feature representation and prediction stability. In data processing, multistation environmental observations are used to construct node inputs, along with missing value handling and normalization strategies to improve model adaptability. For performance evaluation, comparative experiments, hyperparameter sensitivity analysis, and environmental disturbance tests are conducted to assess model performance under different settings. Experimental results show that the proposed method outperforms existing spatiotemporal graph models across multiple metrics, demonstrating strong accuracy, robustness, and structural consistency, and providing an effective solution for high-quality modeling of complex environmental data.