Assessment of Rainfall Data Imputation Techniques: A Comparative Study with Focus on Thai Rainfall Dataset

Addressing missing rainfall data remains a persistent complexity in hydrological modeling, water resource planning, and management. The need for comprehensive and reliable data is critical to allowing efficient water resource planning and management efforts. To address this issue, this study estimates missing values within three distinct datasets characterized by varying degrees of data incompleteness: Set-1 (10% missing data), Set-2 (30% missing data), and Set-3 (50% missing data). Employing various estimation methods, the investigation aimed to identify optimal approaches for accurate imputations. Using data from 31 observation stations in northern Thailand from 1993 to 2022, methods such as arithmetic averaging, long short-term memory recurrent neural network, multiple linear regression, M5 model tree, nonlinear iterative partial least squares, and K-nearest neighbor were evaluated. Among the rainfall stations, Tha Wang Pha and neighboring stations were selected based on Pearson’s correlation. Three datasets with different levels of missing data were created. MLR consistently demonstrated strong performance, particularly in Set-1 with a high R2 (0.94). In Set-2, MLR and LSTM-RNN showed the highest R2 and lower error metrics, with MLR exhibiting an R2 of 0.92 and LSTM-RNN an R2 of 0.91. Despite 30% missing data in Set-2, MLR and LSTM-RNN remained robust. In the most challenging scenario, Set-3 with 50% missing data, LSTM-RNN excelled, exhibiting the highest R2 of 0.90 and low error metrics, benefiting from its ability to capture temporal patterns. These results underline the importance of selecting appropriate models tailored to station characteristics for accurate prediction and monitoring in hydrological studies