A Data-Driven Hybrid Machine Learning and Transfer Learning Algorithm for Long Consecutive Atmospheric Environment Missing Data Imputation

Comprehensive and accurate environmental data form the foundation for studying the correlation between substance properties and the environment. Nevertheless, missing data is inevitable during the data collection process, posing a challenge to research advancement, especially in cases of prolonged and consecutive data gaps. The inherent periodicity of environmental data, which has rarely been considered separately in previous studies, plays an important role in imputing long consecutive missing data. This article combines the advantages of machine learning and transfer learning methods and proposes an imputation algorithm that can fully utilize the inherent periodic components of long consecutive missing data. The algorithm initially utilizes wavelet coherence to assess the periodic coherence of the data, decomposing it into periodic and non-periodic components. Subsequently, the random forest method is employed to impute the periodic components, while the transfer learning approach is used to impute the non-periodic components. By aggregating the imputed values of each component, the final imputation output is obtained. The algorithm was validated using environmental data from five regions in China, demonstrating superior performance compared to existing algorithms across various levels of missing data.