Using Temporal Deep Learning Models to Estimate Daily Snow Water Equivalent over the Rocky Mountains

In the montane western United States, where the majority of downstream water resources are derived from snowmelt, a warming climate threatens the timing and amount of future water availability. It is expected that the fraction of precipitation falling as snow will continue decreasing and the timing of snowmelt will continue shifting earlier in the year with unknown impacts on partitioning between evapotranspiration and streamflow. To assess this, we employ a Snow Storage Index (SSI) to represent the annual temporal phase difference between daily precipitation and daily modeled surface water inputs (SWI, the sum of rainfall and snowmelt), weighted by the respective amounts. We coupled the SSI metric with a Budyko‐based framework to determine the effect of snow water storage on relative hydrologic partitioning across snow‐influenced watersheds in the western U.S. Greater snow water storage was positively correlated with greater hydrologic partitioning to streamflow, particularly in the North Cascades/Cascades (r2: 0.62), Blue Mountains (r2: 0.56), Canadian Rockies (r2: 0.55), Idaho Batholith, (r2: 0.48), and Columbia Mountains/Northern Rockies (r2: 0.45). The weekly SWI:P ratio was an equally strong predictor for hydrologic partitioning, particularly in mid‐spring (e.g., March/April) in the same mountainous areas (r2: 0.62–0.74, across the same eco‐regions). The retention of snow water storage and subsequent release of stored water in summer months resulted in increased hydrologic partitioning to streamflow. If SSI decreases with future warming, the volume of water partitioned streamflow will decrease non‐uniformly across the western U.S. with substantial implications for ecosystems and agricultural, industrial, and domestic water supplies.

Section: Methodsmentioning

confidence: 99%

Effects of Snow Water Storage on Hydrologic Partitioning Across the Mountainous, Western United States

Hale

Musselman

Newman

et al. 2023

“…Our predictions for SNOTEL stations, extrapolation over the Rocky Mountains along with the necessary code can be accessed at S. Duan et al. (2022). SNOTEL SWE observations can be accessed at https://data.nal.usda.gov/dataset/snowpack-telemetry-network-snotel (USDA Natural Resources Conservation Service, 2022) and https://www.pnnl.gov/data-products.…”

Section: Data Availability Statementmentioning

confidence: 99%

Using Temporal Deep Learning Models to Estimate Daily Snow Water Equivalent Over the Rocky Mountains

Duan,

Ullrich,

Risser

et al. 2024

In this study we construct and compare three different deep learning (DL) models for estimating daily snow water equivalent (SWE) from high‐resolution gridded meteorological fields over the Rocky Mountain region. To train the DL models, Snow Telemetry (SNOTEL) station‐based SWE observations are used as the prediction target. All DL models produce higher median Nash‐Sutcliffe Efficiency (NSE) values than a conceptual SWE model and interpolated gridded data sets, although mean squared errors also tend to be higher. Sensitivity of the SWE prediction to the model's input variables is analyzed using an explainable artificial intelligence (XAI) method, yielding insight into the physical relationships learned by the models. This method reveals the dominant role precipitation and temperature play in snowpack dynamics. In applying our models to estimate SWE throughout the Rocky Mountains, an extrapolation problem arises since the statistical properties of SWE (e.g., annual maximum) and geographical properties of individual grid points (e.g., elevation) differ from the training data. This problem is solved by normalizing the SWE with its historical maximum value to alleviate extrapolation for all tested DL models. Our work shows that the DL models are promising tools for estimating SWE, and sufficiently capture relevant physical relationships to make them useful for spatial and temporal extrapolation of SWE values.

“…That early work also explored the use of simple recurrent neural networks (RNNs) to emulate system memory and state dynamics (Hsu et al, 1997). Recently, however, with the development of gated RNN's, and in particular the long-short-term memory (LSTM) network (Hochreiter & Schmidhuber, 1997), the ability of ML to advance the modeling of dynamical hydrological processes has been dramatically demonstrated, not only for catchment-scale RR modeling (Arsenault et al, 2023;Feng et al, 2020;Kratzert et al, 2018Kratzert et al, , 2019Lees et al, 2021), but also for snowpack modeling (Duan et al, 2023;Wang et al, 2022), and in many other contexts (Than et al, 2021;Zhi et al, 2023) that are relevant to water resource management, such as addressing the potential impacts of changing climate (Sungmin et al, 2020). However, concerns have been raised about the physical interpretability of ML-based models, and considerable attention is now being devoted to addressing this issue (Guidotti et al, 2018;Molnar, 2022;Molnar et al, 2020;Montavon et al, 2018;Samek et al, 2019); see also Fleming et al (2021) and McGovern et al (2019) in the hydrological and meteorological contexts respectively.…”

Section: Ml-based Modeling Of the Rainfall-runoff Modeling Systemmentioning

confidence: 99%

“…That early work also explored the use of simple recurrent neural networks (RNNs) to emulate system memory and state dynamics (Hsu et al., 1997). Recently, however, with the development of gated RNN's, and in particular the long‐short‐term memory (LSTM) network (Hochreiter & Schmidhuber, 1997), the ability of ML to advance the modeling of dynamical hydrological processes has been dramatically demonstrated, not only for catchment‐scale RR modeling (Arsenault et al., 2023; Feng et al., 2020; Kratzert et al., 2018, 2019; Lees et al., 2021), but also for snowpack modeling (Duan et al., 2023; Wang et al., 2022), and in many other contexts (Than et al., 2021; Zhi et al., 2023) that are relevant to water resource management, such as addressing the potential impacts of changing climate (Sungmin et al., 2020).…”

Section: Introduction and Scopementioning

confidence: 99%

A Mass‐Conserving‐Perceptron for Machine‐Learning‐Based Modeling of Geoscientific Systems

Wang,

Gupta

2024

Although decades of effort have been devoted to building Physical‐Conceptual (PC) models for predicting the time‐series evolution of geoscientific systems, recent work shows that Machine Learning (ML) based Gated Recurrent Neural Network technology can be used to develop models that are much more accurate. However, the difficulty of extracting physical understanding from ML‐based models complicates their utility for enhancing scientific knowledge regarding system structure and function. Here, we propose a physically interpretable Mass‐Conserving‐Perceptron (MCP) as a way to bridge the gap between PC‐based and ML‐based modeling approaches. The MCP exploits the inherent isomorphism between the directed graph structures underlying both PC models and GRNNs to explicitly represent the mass‐conserving nature of physical processes while enabling the functional nature of such processes to be directly learned (in an interpretable manner) from available data using off‐the‐shelf ML technology. As a proof of concept, we investigate the functional expressivity (capacity) of the MCP, explore its ability to parsimoniously represent the rainfall‐runoff (RR) dynamics of the Leaf River Basin, and demonstrate its utility for scientific hypothesis testing. To conclude, we discuss extensions of the concept to enable ML‐based physical‐conceptual representation of the coupled nature of mass‐energy‐information flows through geoscientific systems.