Spatiotemporal techniques for imputing sequential missing values in rainfall data collected from Ratnapura area: A comparative study

Modeling and forecasting rainfall is vital for issuing early flood warnings, but missing data, particularly sequential gaps, make this task challenging. Even though there are several techniques proposed in the literature aiming at the estimation of sequential missing values, selecting the most suitable technique is challenging. It is worth considering both temporal patterns and spatial patterns in such imputations. This study evaluates four widely used spatiotemporal techniques, Knearest neighbors (KNN), MissForest, denoising autoencoder, and spatiotemporal kriging, to identify the best approach for imputing sequential missing values of rainfall at Ratnapura. Analysis was carried out using daily rainfall data of Rathnapura and six neighboring gauging stations, collected from 2015 to 2019. The accuracy of each estimated series was evaluated against the actual data using root mean square error (RMSE) and mean absolute error (MAE). The results show that the KNN is the most suitable technique under both wet and dry seasons, showing lower RMSE and MAE values. MissForest also performed well, surpassing denoising autoencoder and spatiotemporal kriging, though it generally fell slightly short of KNN. Among the methods compared, spatiotemporal kriging demonstrated the lowest accuracy in both wet and dry seasons due to a lack of reference stations and other meteorological variables. However, none of the techniques were successful in accurately imputing sudden extreme rainfall events, highlighting an important limitation. Forecasted rainfall at the Ratnapura station using both the imputed data and the actual data revealed that the forecasts obtained using the imputed data were not significantly different from those obtained using the actual data.