Imputation of Missing Values for Generating Rainfall Data Using a Multi-Method Approach

The daily rainfall records of Cibeureum Station (West Java, Indonesia) have 2,172 missing values which amounts to 23.8 % of the data from 1999 to 2023 and is the most prevalent after the year 2010. These gaps undermine the effectiveness of flood forecasting and long-term trend analysis. To evaluate the effectiveness of various gap-filling techniques, we applied the three patterns of MCAR (10 percent random), MAR (weekly blocks), and MNAR (30 percent during light rain days, 5 percent during heavy rain days) to the complete segment of 1999-2010. Five methods were compared: mean value, linear interpolation, SARIMA, KNN, and a stacked LSTM autoencoder (three encoder and one decoder LSTM layers) with MAE, RMSE, NSE, KGE, and MBE. The best results in RMSE of 12.7 mm and positivity for NSE of $0.03-0.29$ in all scenarios was observed in the LSTM method. Linear interpolation performed nearly as well for isolated random gaps. The mean value method was the worst performer across the board. These findings demonstrate that even in sparse tropical datasets, deep sequence models can preserve the daily extremes and keep annual bias within $\pm 1 \%$. Future work will test hybrid models (e.g., SARIMA-LSTM, MissForest-LSTM). In addition, we will extend the framework to neighbouring stations-Bandung, Soreang, Pangalengan-to assess spatial transferability.