Forecasting of Reservoir Water Level by Remote Sensing and Deep Learning

Abstract

Abstract Accurate reservoir water level prediction is crucial for the safe operation of reservoirs and the utilization of their functions. Traditional physical-based water level forecasting methods rely heavily on auxiliary data, such as precipitation and reservoir outflow. However, obtaining timely and reliable data on reservoir discharge flow can be expensive and impractical due to limitations in infrastructure or data accessibility. To address this issue, the current study utilized multi-source remote sensing data to extract a time series of reservoir storage volume. A technical framework for predicting water levels using deep learning models and remote sensing technology was proposed. To validate the effectiveness of this method, we compared the predictive accuracy of reservoir water levels among 16 different machine learning input scenarios. The results indicate that the model incorporating water level, rainfall, water surface area, and daily changes in reservoir storage volume as input data performed the best. Compared to the input data that did not consider water surface area and daily changes in reservoir storage, it demonstrated higher accuracy, with an increase in R2 value by 1.13%, RMSE increased by 52.17%, and MAE increased by 63.80%. The framework proposed in this study offers a reliable method for predicting reservoir water levels in the absence of operational data for reservoirs.