Construction of a near-real-time long-term monthly water body dataset by fusing JRC MWH and Sentinel data and its application in reservoir dynamics monitoring

The Joint Research Centre Monthly Water History (JRC MWH) dataset has been widely applied in global water resource studies but suffers from data gaps, anomalies, and poor timeliness. This study constructs a Near-Real-Time Long-Term Monthly Water Body Dataset (NMWBD) by fusing JRC MWH and Sentinel data, and applies it to reservoir monitoring in the Fenshui River Basin, southeastern China. Data gaps are processed with preliminary gap-filling using dynamic thresholds method, followed by fine-tuning gap-filling with a deep learning model which combines Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM). To address data anomalies, a sliding-window anomaly detection method is used, which are then corrected using adjacent normal months. To improve the timeliness of JRC MWH, Sentinel-1 and Sentinel-2 images are used to extract recent monthly water bodies with a threshold-based segmentation algorithm, using the SDWI and MNDWI indices in combination with Otsu’s method. Finally, the NMWBD, covering the period from 1990 to 2023, is constructed by concatenating processed historical JRC MWH data (1990–2021) with recent water body data extracted from Sentinel images (2019–2023), with the overlapping years (2019–2021) directly replaced by the latter. The dataset’s accuracy is validated through visually collected water body samples from Google Earth images, with an average Kappa coefficient of 0.85. Comparisons between ground-observed reservoir water levels and extracted water surface areas of reservoirs also demonstrate high consistency, with a Pearson correlation coefficient exceeding 0.85 and a coefficient of determination above 0.8. NMWBD is employed to identify reservoirs in Fenshui River Basin. Reservoirs larger than 0.1 km² are fully identified, while smaller ones achieve a true positive rate of 0.91. The variations in extracted water surface area accurately reflect the long-term trends of reservoirs in the basin, and capture annual and seasonal fluctuations driven by both precipitation variations and reservoir operations.