Benchmarking Singular Spectrum Analysis for Imputation in the Sudanian Zone with Validation Samples

Abstract

This paper explores the performance of various imputation techniques for climatic time series data acquired in the Sudanian Zone, West Africa. We evaluated three methods: Singular Spectrum Analysis (SSA) and its multivariate counterpart (mSSA), as well as Random Forest and Multiple Linear Regression models, both implemented with the Multiple Imputation by Chained Equations (MICE) strategy. Results show that SSA is superior in capturing patterns for independent variables, while mSSA excels with highly correlated features. Although mSSA processes data faster, SSA remains preferable overall and consistently outperformed other models when most sensors failed to record data simultaneously. Challenges with lengthy missing segments impacted imputation accuracy. Our study advances imputation techniques, providing insights into handling missing data in climatic time series analysis.