Year of Award

2024

Document Type

Thesis

Degree Type

Master of Science (MS)

Degree Name

Geosciences

Other Degree Name/Area of Focus

Glaciology / Computer Science

Department or School/College

Geosciences

Committee Chair

Joel T. Harper

Commitee Members

Javier Pérez Álvaro, Jesse V. Johnson, W. Payton Gardner

Keywords

snowpack water AI machine learning statistics

Subject Categories

Artificial Intelligence and Robotics | Data Science | Glaciology | Hydrology | Numerical Analysis and Scientific Computing

Abstract

The timing of water release from the snowpack plays key roles in ecosystem services, groundwater recharge, and water resource management. However, two internal barriers in a standing snowpack must be overcome before runoff can outflow from the base: 1) the cold content must be exhausted, and 2) the interconnected network of snow grains must be filled with liquid water to residual saturation. Expressing the liquid water as latent heat allows the two barriers to be grouped as an energy (J/m²) to define a snowpack’s Runoff Energy Hurdle (REH). The growth and loss of REH is driven by evolution of pore space and heat exchange with the atmosphere, both of which are modulated by the seasonal cycle. The snowpack is commonly referred to as ‘ripe’ after energy input to the snowpack from heat/water has exhausted REH. Here, we define activation as the switch to REH=0 of a given mass of snow, at a location and time, indicating that the snowpack is capable of runoff outflow from the base. The spring activation is defined as the highest magnitude of snow water equivalent activated in a water year. First, we seek to quantify the spatial and temporal variability of activations and spring activations within the snowpack of the South Fork of the Flathead River, Montana, a 4341 km² mountainous watershed. We use the 2005-2023 daily outputs of the SNODAS model to compute distributed REH and activations. For north aspects, spring activation occurs 3.2 days later per 100 m elevation gain, and for southwest aspects it is 4.8 days later per 100 m elevation gain. We classify yearly REH time series in 4 clusters using K-Means. Cluster 1 contains the lowest REH locations, always including the valley floor but extending to higher elevations during warm and low snow accumulation years. During low snow years, up to 92% of the watershed is in cluster 1 and the snowpack is activated numerous times throughout the winter season. In contrast, cluster 2 is represented all years and cluster 3 15/19 years, they characterize a snowpack with greater REH having 69% of the basin’s snow water equivalent volume, sitting at higher elevations and ripening within a two month period. Finally, cluster 4 only appears during exceptionally high-accumulation years, drawing the locations with the greatest snow accumulation from cluster 3 and activating within a one-month period. Second, we assess predictive models for spring activation timing for clusters 2 through 4, based on partial REH time series until 1 March. We find that K-Means clustering yields the best predictions of spring activation timing for the high-REH clusters (3 and 4) that can only ripen once energy inputs become great in the spring; the two respective mean absolute errors (MAE) are 12.3 and 8.8 days, which is an improvement from the historical mean of 16.9 days. A Long Short-Term Memory (LSTM) neural network is deployed, taking advantage of the temporal correlation and embracing the non-linearity of REH, but it only outperforms K-Means (MAE=21.3 days)for cluster 2 with a mean absolute error of 19.3 days, barely outperforming the cluster 2’s historical mean. Years mostly classified in cluster 1 are expected to become more frequent in a warmer world, bringing numerous activations throughout the winter that cannot be predicted, unlike higher clusters’ activations, therefore reducing the predictability of spring activation timing for flood forecasters and dam managers. Conversely, increased water vapor in the atmosphere with the rising temperatures will likely promote continued development of cluster 4 in heavy accumulation years, which has a narrow window for spring activation timing.

Download

Share

COinS
 

© Copyright 2024 Clement Cherblanc