UC Berkeley

Monitoring the snowpack is crucial for water management, flood control and hydropower optimization. Traditional regression methods often result in low accuracy runoff predictions.Existing ground-based real-time measurement systems are in majority installed at low elevations with poor physiographic representation. This thesis presents a system for better Snow Water Equivalent (SWE) and runoff estimation. The autonomous end-to-end Wireless Sensor Network (WSN) that leverages the Internet of Things (IoT) technology provides mountain hydrology measurements in near real-time. At its core lies an ultra-low power, radio channel-hoping, and self-organizing mesh secured with a rugged weather-sealed design, data replication and remote network health monitoring. Three WSNs are installed throughout the North Fork of the Feather River in Northern California upstream of the Oroville dam. Elevation, aspect, slope and vegetation determine network locations. Data show considerable spatial variability of snow depth, and that existing operational autonomous systems are non-representative spatially, with biases reaching up to 50%. Combined with existing systems, WSNs better detect precipitation timing and phase, monitor sub-daily dynamics of infiltration and surface runoff, and inform hydro power managers about actual ablation and end-of-season date across the landscape. A wet and dry year exhibit strong multi-scale inter-year spatial stationarity with major rank conservation. Elastic Net regression shows that dominant features at the sub-km2 scale are site-dependent and differ from the watershed scale. Based on the Nearest Neighbor (NN) with a Landsat assimilated historical product, explanatory variables consistently explain up to 90% of the variance in the watershed-scale SWE for both years. Lagged cross correlation of snowmelt with stream flow measurements show improvement of up to 100% compared to existing systems. Ensemble Optimal Interpolation (EnOI) update of background SWE fields from Landsat and LiDAR products provide accurate high resolution estimates of spatial SWE for areas with parsimonious sensors. Results show a minimum RMSE of 22% and 30% at 90 m and 50 m resolutions respectively. Compared with SNODAS, reduction in error is up to 55% and 80%, with LiDAR as reference.