|
Estimating Water-Surface Elevation in the Upper Gotvand Dam Reservoir by Integrating Remote Sensing and Hydrometeorological Variables within a Machine-Learning Framework
|
Javad Zahiri *   |
| Agricultural Sciences and Natural Resources University of Khuzestan |
|
|
Abstract: (8 Views) |
| Reservoir water-surface elevation is one of the most important variables required for water-resources monitoring and management; however, continuous access to in situ observations is constrained by technical and operational limitations in many reservoirs. This study aimed to develop a remote-sensing and data-driven framework for estimating the water-surface elevation of the Upper Gotvand Dam Reservoir and to investigate the applicability of spectral, elevation, and hydrometeorological variables in this process. Daily reservoir water-level data covering the May 2016 to May 2026 were integrated with Sentinel-2 imagery, the ALOS World 3D–30m digital surface model (AW3D30 DSM), and ERA5-Land data. First, the normalized difference water index (NDWI) was calculated from Sentinel-2 images, and five thresholds of 0.00, 0.05, 0.10, 0.15, and 0.20 were evaluated for extracting the reservoir’s water boundary. The preliminary water-surface elevation was then estimated by overlaying the extracted boundary on the digital surface model and calculating the median elevation of the boundary pixels. In addition, precipitation, temperature, evaporation, runoff, snow-water equivalent, and snowfall variables were extracted for the HydroBASINS-based upstream sub-basins and used in the modeling process by generating accumulated and averaged temporal variables. Four models—XGBoost, a deep neural network (DNN), a convolutional neural network (CNN), and a long short-term memory network (LSTM)—were employed to compare the performance of the data-driven approaches. Evaluation of the NDWI thresholds showed that 0.00 provided the best performance and that increasing the threshold resulted in a continuous decline in water-level estimation accuracy. Among the data-driven models, XGBoost achieved the best performance on the test dataset, with R² = 0.98, RMSE = 1.07 m, and NSE = 0.98. SHAP analysis also showed that the preliminary water-level variable derived from the satellite imagery and digital surface model was the most important predictor, while the hydroclimatic variables played a complementary role in improving the estimates. Comparison of the model results using a Taylor diagram further emphasized the greater stability of XGBoost and its superior ability to reconstruct the magnitude and pattern of water-level variations in the test data. Overall, the results demonstrated that integrating remote-sensing information, topographic data, and upstream hydrometeorological conditions with data-driven algorithms, particularly XGBoost, can provide an effective approach for estimating the water-surface elevation of the Upper Gotvand Dam Reservoir. |
|
Article number: 6 |
| Keywords: Water level, Upper Gotvand Dam, Sentinel-2, NDWI, AW3D30 DSM, ERA5-Land, XGBoost, Deep learning. |
|
|
|
Type of Study: Research |
Subject:
Special
Received: 2026/09/2 | Revised: 2026/09/29 | Accepted: 2026/10/12 | ePublished ahead of print: 2026/10/15
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
