Enhancing flow-through aquaculture system monitoring: A comparative study of machine learning algorithms for missing-data imputation

This study evaluated flow-through aquaculture systems by assessing the efficiency of various machine learning algorithms for imputing missing water-quality data, including dissolved oxygen, water temperature, pH, and salinity. Artificial missing data were generated based on real-world missing data mechanisms, and a comprehensive statistical analysis of the data characteristics was conducted to identify suitable imputation methods. Results showed that basic imputation methods like linear interpolation, often insufficient for datasets with high variability and non-linear relationships, performed well for certain data distributions, particularly for salinity and pH data with high kurtosis and symmetric distributions. However, advanced machine learning-based imputation techniques, especially TimesNet, consistently outperformed other methods in handling complex and variable patterns in the water-quality data. This study underscores the importance of selecting appropriate imputation methods based on data properties to enhance environmental monitoring systems in aquaculture and improve operational efficiency and sustainability. • Machine learning algorithms' performance in imputing missing water-quality data in an aquaculture system was evaluated • Artificial missing data were generated based on actual missing data mechanisms • Linear interpolation suited simpler data distributions, but TimesNet excelled in complex patterns • Imputation methods should be selected based on the statistical properties of the data