Estimating Reservoir Sedimentation Using Deep Learning and The Usace Rsi System Dataset

Many reservoirs across the nation are becoming progressively filled with sediment, which reduces their effectiveness and increases the cost of preventing water intakes and outlets from being buried. The US Army Corps of Engineers (USACE) developed the Reservoir Sedimentation Information (RSI) system to assess reservoir aggradation and track the dam operation suitability for water-resource management, protecting recreational areas, and ensuring dam safety. This unprecedented dataset contains information on approximately 400 dams (excluding navigation structures). Within the database, 184 reservoirs include three or more surveys with elevation-capacity and/or elevation-surface area curves. Given that over 90,000 dams exist in the US, the RSI dataset represents less than 1% of the US dams. Thus, there is a critical need to develop methods for estimating reservoir sedimentation for unmonitored sites. The goal of this research was to create a universal method for estimating reservoir sedimentation rates using RSI data and other supplemental resources which can then be used to estimate current and forecast future conditions of reservoirs within the contiguous United States. To meet this objective, geospatial tools were utilized to build a composite dataset that complemented the data available at each reservoir within the RSI system. These parameters included precipitation and watershed characteristics. Initially, a feature correlation analysis enabled the refinement of the composite dataset. Nine models were then used on the compiled dataset to determine its accuracy at predicting capacity loss for the RSI system. These models included four supervised machine learning models, four deep neural network (DNN) models, and a multilinear power regression model. Of these nine models, a DNN model, which contained a progressively increasing node and layer construction, was deemed the most accurate, with R2 values from its calibration and validation datasets being 0.83 and 0.70, respectively. This best model was then recalibrated over the entire dataset, which showed greater capacity loss predictive accuracy for the RSI system, with an R2 of 0.81. The prediction models can be used to evaluate the capacity loss of unmonitored reservoirs, identify reservoirs with the highest risk of losing functionality, and forecast future capacity loss under different climate change scenarios.