Missing Values Imputation for Spatio-Temporal Climate Variable using Graph Convolutional Networks (GCN)

Abstract

Climatic time series analysis often suffers from an issue of missing values. Which directly affects the accuracy of analysis and modeling especially with multi-site data which include spatial and temporal dependencies. Although traditional imputation methods do exist, most of them depend mainly on temporal information which limits its efficiency especially with consecutive missing values or when the temporal data are weak. This limitation becomes more critical under the Block Missing Pattern, where consecutive missing values reduce the availability of temporal information and increase the difficulty of the imputation process. This study applies a spatially structured Graph Convolutional Network (GCN) model for missing-value imputation in multivariate climatic time series collected from multiple meteorological stations. covering the period from 1994 to 2024 with a total of 11,329 daily observations. To reduce seasonal heterogeneity, the data were first stratified into hot and cold seasons. Each seasonal dataset was then divided into training (80\%), validation (10\%), and testing (10\%) subsets for model development and performance evaluation. Overall, this study confirms that the use of models that depend on spatial structures is an effective method for handling missing value in time series, and paves the way for developing more advanced models that combines between the spatial and temporal dimensions. The model’s performance has been evaluated under the Block Missing Pattern in both of the hot and cold seasons using RMSE, MAE and MAPE metrics, as well as comparing the performance with K-Nearest Neighbors (KNN), Linear Interpolation (LI), and GRU-D. The results show that the Graph Convolutional Network provides a more efficient representation of the climatic data, especially in cases with weak or unavailable temporal data, where a higher accuracy can be achieved by using spatial relationships between stations.{The proposed model achieved the lowest error values across most climatic variables in both seasons} The results also indicate that the nature of the data plays an essential role in the performance, where the decrease of variability and the stability of values in the hot season contributed to the improvement of the missing data imputation.