Bridging data gaps in streamflow records: a comparative study of LSTM and interpolation methods

Reliable daily streamflow data are essential for effective water resource management, particularly for flood prediction and for assessing the impacts of climate change and human activities on future water supplies. In data-scarce regions such as the Pra River Basin, limited streamflow records hinder understanding of long-term trends and lead to inaccurate forecasts, which may have serious consequences. This study evaluates three approaches for filling gaps of varying lengths in daily streamflow data: linear interpolation (LI), which estimates missing values based on adjacent days, cubic spline interpolation (SI), which models smooth variations in data, and the deep learning long short-term memory network (LSTM), which captures complex temporal dependencies. These methods were selected for their simplicity, low data requirements, and computational efficiency. Four synthetic gap scenarios of varying durations were tested, and their performance was assessed using the Coefficient of Determination (R²), Nash–Sutcliffe Efficiency (NSE), Percentage Bias (PBIAS), and Root Mean Squared Deviation Ratio (RSR), all under consistent gap settings. Results indicate that interpolation techniques (LI and SI) can reliably reconstruct gaps of one to seven days, whether sequential or random, due to the smooth short-term variations in streamflow. In contrast, the LSTM network, which captures long-term dependencies, outperformed interpolation methods when reconstructing combinations of different gap lengths in the daily streamflow data up to a year in advance. However, gaps longer than one year could not be effectively addressed, reflecting the inherent limitations of daily streamflow data. Overall, this study provides practical guidance for selecting appropriate gap-filling methods in hydrological datasets, offering both methodological insights and applied value for water resource management in data-scarce basins.