A hybrid satellite data‑machine learning framework for Dez Dam water level monitoring

Introduction
Surface water resources in the form of lakes, rivers, reservoirs, wetlands, snow, and glaciers play an essential role in various aspects of life on Earth. These valuable resources influence ecosystems, hydrology, and climate while providing numerous benefits, including urban, agricultural, and industrial water supply, supporting wildlife and fisheries, and creating recreational opportunities for humans. Among these, reservoirs regardless of their size are of great importance in the integrated management of water resources in watersheds. Large and medium reservoirs play a vital role in flood control and hazard mitigation, water supply and irrigation, hydropower generation, and runoff regulation. Small reservoirs, despite their large numbers, play an irreplaceable role in ensuring drinking water and food security, aquaculture, and water resource allocation. Reservoir storage capacity is an important indicator of its healthy functioning, which is closely related to regional climate; therefore, regular and accurate monitoring of reservoir water volume is essential for the protection and rational exploitation of water resources and the formulation of related policies. This dynamic monitoring is particularly important in arid regions for water resource assessment, hydropower generation, and irrigation. However, existing methods for calculating reservoir water levels are mainly based on field measurements, which limits their application in data-sparse regions. Unlike previous studies, the present study provides a comprehensive framework for monitoring the water level of the Dez Dam reservoir using Sentinel-2 imagery and the SRTM digital elevation model within the Google Earth Engine platform. In this framework, the support vector machine model, uncertainty analysis, and sensitivity analysis are all implemented in an integrated framework.
 
Materials and Methods
In this study, water level variations of the Dez Dam reservoir were analyzed using satellite data from 2018 to 2024. Sentinel-2 surface reflectance products from the COPERNICUS collection were utilized, employing bands B3, B8, and B11. The Normalized Difference Water Index (NDWI) and Modified Normalized Difference Water Index (MNDWI) were calculated to separate water surface from surrounding dry land. Water body boundaries were extracted, and a one-pixel buffer around the water boundary was created to extract elevation values from the SRTM digital elevation model (30 m resolution), with the mean elevation along this boundary considered as the reservoir water level. For each observational date, five input variables were extracted: satellite-derived water level, mean NDWI, mean MNDWI, reservoir surface area, and the 5‑day lagged observed water level. An ensemble-based support vector machine regression model (SVM) with ten independent models was developed using random sampling of the training dataset, and the average prediction of all models was considered as the final output. The ensemble approach, by averaging ten independently trained models, effectively reduced variance and improved generalization compared to a single SVM model. For uncertainty estimation, conformal prediction was applied without assuming any specific statistical distribution for the errors. This method generated prediction intervals by calculating nonconformity scores on a calibration set, and then deriving a threshold at the 90th percentile. The resulting intervals were designed to provide distribution-free uncertainty bounds with a target 90% coverage for new observations. Permutation importance was used for sensitivity analysis to quantify the contribution of each input parameter to the model's predictive performance.
 
Results and Discussion
The SVM model was evaluated under two scenarios. In the first scenario, four satellite-derived variables, including satellite-derived water level, reservoir surface area, NDWI, and MNDWI, were used as model inputs. This scenario yielded RMSE values of 2.29 m for training and 2.64 m for validation, with corresponding (R²) values of 0.96 and 0.94, respectively. Error analysis showed slight overestimation at low water levels (below 320 m) and underestimation at high water levels (above 345 m), indicating reduced accuracy in reproducing extreme hydrological conditions such as drought and flood peaks. In the second scenario, the 5-day lagged water level was added to the four satellite-derived inputs, resulting in improved model performance. RMSE decreased to 1.36 m for training and 1.97 m for validation. Although validation R² and NSE remained nearly unchanged between the two scenarios, the lower RMSE, and narrower prediction intervals indicate a more accurate and stable model. Uncertainty analysis using conformal prediction further showed that the prediction interval width decreased from 6.81 m in the first scenario to 4.24 m in the second scenario, although conformal coverage remained slightly below the target 90% in both scenarios (84% and 85%, respectively). These results indicate that incorporating antecedent water-level information substantially improves both predictive accuracy and uncertainty performance of the SVM model. Permutation importance analysis further showed that the 5-day lagged water level was the most influential predictor (ΔRMSE = 4.05 m), followed by satellite-derived water level (2.85 m) and reservoir surface area (0.94 m), whereas NDWI and MNDWI exhibited comparatively minor direct contributions to model performance.
 
Conclusion
This study developed an integrated framework within Google Earth Engine to estimate the water level of the Dez Dam reservoir from 2018 to 2024 using Sentinel-2 imagery, SRTM digital elevation data, and an SVM model. Two modeling scenarios were evaluated. In the first scenario, four satellite-derived variables, were used as model inputs, yielding a validation RMSE of 2.64 m. In the second scenario, the 5-day lagged water level was added to these variables, improving model performance and reducing the validation RMSE to 1.97 m. Uncertainty analysis using conformal prediction showed that the prediction interval width decreased from 6.81 m in the first scenario to 4.24 m in the second scenario, indicating reduced predictive uncertainty when antecedent water-level information was incorporated. Nevertheless, because the first scenario relies solely on remotely sensed variables and does not require ground-based observational data, it provides a practical solution for reservoir monitoring in ungauged or data-scarce regions, with lower accuracy than the hybrid scenario. Overall, the proposed framework offers a reproducible and automated approach for water-level estimation, while future studies may further improve performance by incorporating additional hydrological variables such as inflow, outflow, and evaporation, as well as radar observations (e.g., Sentinel-1) to enhance temporal coverage under cloudy conditions.