Exploring the Integration of Process-Based and Deep Learning Approaches for Modeling Snowpack Dynamics
Abstract
Given the global importance of snow as a source of freshwater, accurate estimation of the spatio-temporal distribution of snow water equivalent (SWE) is essential. Snow-rain partitioning and melt are the two fundamental processes most frequently being encoded into a process-based model of land-surface snow dynamics. While the standard scientific approach is to progressively incorporate and test various hypotheses regarding appropriate model structure, recent advances in machine/deep learning (ML/DL) raise the possibility that data-based methods can provide the complementary ability to discover and exploit relevant information from multiple data sources, thereby learning a better representation of the underlying data generating process while achieving improved predictive performance. To date, however, the predictive success of ML/DL methods has not necessarily translated into significantly improved knowledge of the processes underlying snow dynamics. Here, we explore the possibility of enhanced learning by integrating process-based and data-based strategies for building a representation of land-surface snow dynamics. As benchmark representational strategies, we use the process-based SNOW17 model, and the ML-based LSTM network developed by Wang et al (2021). The dataset includes the University of Arizona (UA) 4-km daily ground-based SWE, meteorological forcing from PRISM (precipitation, mean temperature, dew point temperature and vapor pressure deficit), and other overarching ancillary static features. Our investigation includes: 1) application of spatial regularization constraints to the parameters of SNOW17, 2) development of a parsimonious process-based modeling system that uses symbolic regression to construct physically understandable process representations, and 3) use of different types and sequence lengths of input information to construct the gating operations that enable the LSTM network to respond appropriately to varying hydro-geo-climatic context. Our ultimate goal is to develop a hybrid modeling system with sufficient predictive accuracy and robustness, while seeking a parsimonious representation of the underlying data generating process.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35S1257G