Enhanced tensor factorization for spatiotemporal imputation of high-frequency water quality monitoring data

Abstract

Automatic high-frequency monitoring (AHFM) of water quality is increasingly being deployed to enhance the management and scientific understanding of eutrophic lakes. However, the prevalence of missing data in such systems poses significant challenges, potentially compromising decision-making processes and distorting analytical or modelling outcomes. This study proposes an enhanced tensor factorization model, termed the Diverse Biases-integrated Adaptive Latent-factorization-of-tensors (DBAL), which decomposes high-dimensional data into low-rank components while incorporating diverse linear biases to capture temporal fluctuations and employing a differential evolution algorithm for adaptive hyperparameter optimization. Extensive empirical validation using real-world AHFM datasets from a large eutrophic lake in China demonstrates that our proposed DBAL consistently outperforms the state-of-the-art models, achieving 5.6 %–50.3 % and 5.7 %–43.7 % reductions in RMSE and MAE, alongside a 0.3 %–6.1 % improvement in R². Notably, DBAL demonstrates variable-specific performance, with particularly accurate imputation for stable physical parameters while revealing greater challenges for biologically-active variables that exhibit stronger temporal dynamics.