Multivariate Analysis and Anomaly Detection of a US Reservoir Sedimentation Data Set

Sedimentation processes in reservoirs can jeopardize their functionality and compromise dam safety. Climate change and associated hydrologic uncertainty are introducing additional stressors to US reservoirs, and data-driven indicators of climate impacts on upstream soil erosion and reservoir sedimentation processes are crucial to evaluate their aggradation and life expectancy. The US Army Corps of Engineers developed the Enhancing Reservoir Sedimentation Information for Climate Preparedness and Resilience (RSI) system to consolidate historical information of elevation-capacity surveys. However, the multiple surveying technologies, protocols, and computational analysis methods used over the service life of reservoirs can impact the quality of reservoir survey data in the RSI system. The objective of this study was to develop a methodology to detect anomalous records and identify multivariate relationships between historical sedimentation data for 184 US reservoirs and associated watershed variables. For this purpose, unsupervised machine learning techniques including principal component analysis (PCA), autonomous anomaly detection, and Kolmogorov–Smirnov and Efron anomaly detection were assembled in an anomaly detection protocol that led to the detection of 20 reservoirs with anomalous records. The variables contributing most to anomaly detection were related to elevation characteristics (watershed and channel slopes, and minimum elevation), precipitation (maximum and cumulative monthly precipitation), dam properties (time since dam completion and initial trap efficiency), and curve number (CN). PCA results indicated that reservoirs in the Mediterranean California ecoregion, although experiencing substantial extreme precipitation events, had small basin areas and CN values that reflected in small capacity losses, contrasting with larger capacity losses found at reservoirs in the Great Plains and Eastern Temperate Forests ecoregions. The developed anomaly detection protocol represents a powerful tool for the analysis and monitoring of this large and heterogeneous data set with the potential of providing reliable information on the impacts of historical climate and watershed properties on erosion and sedimentation processes in US reservoirs.

The US Army Corps of Engineers created the Reservoir Sedimentation Information (RSI) system to compile historical reservoir elevation-capacity data collected using various measurement protocols, instruments, and analysis methods. These differences in data collection and analysis methods, in addition to any human error, can result in anomalies that require detection and correction before the dissemination of the data set for further usage. Data anomalies are values that deviate from normal or expected patterns. Apparent erroneous data, related to duplicate records or increases in reservoir capacities, can be flagged through a preliminary analysis. However, the detection of anomalies in an automated and fully data-driven way represents a powerful tool for the maintenance and monitoring of this large and heterogeneous data set. A depurated RSI data set is a potential major data source for large-scale and long-term studies related to sedimentation rates and suspended solid loads in freshwater systems due to the spatial and temporal scale of its records. This kind of data set will allow the development of effective management plans for reservoir operation, maintenance, and upstream erosion control as well as enabling the indirect monitoring of suspended sediment loads in freshwater systems at a nationwide scale.