This study applies Long Short-Term Memory (LSTM) Networks to investigate different regionalization strategies regarding snow accumulation and melt, using the University of Arizona (UA) 4-km ground-based daily snow dataset over the CONUS. As a physically based benchmark we use the Snow-17 model.
First, a location-agnostic LSTM is pre-trained to learn the general underlying structure of the dynamical relationship between dynamical forcing and snow water equivalent (SWE) using the PRISM (mean/dew point temperature, precipitation, and vapor pressure deficit) and NLDAS2 (longwave/shortwave radiation) datasets. Next, based on the spatial proximity assumption, we impose different kinds of regularization strategies over LSTMs by proposing a CONUS training and a regional training strategy where the network inputs are augmented with a pre-selected important static feature and the physiographical region ID. We demonstrate the model skills (each KGE component and NSE) over CONUS thereby providing insight on how different spatial regions can share a common dynamical model structure with different parameter values.
Lastly, we show the grouping results of 0.3 million snowy pixels obtained from k-meaning clustering using 1) meteorological variables, 2) ancillary variables (land cover type, vegetation, physiographic, wind speed, cloud cover and snow signatures), and 3) both meteorological and ancillary variables as input features thereby demonstrating how to extend the current regionalization ideas to the concept in terms of grouped/hydrologic response unit (GRU/HRU).
In the future, we will 1) suggest the optimal number of regional LSTM models used for the simulation over CONUS, 2) compare the predictive performance against the results obtained via spatial proximity assumption over the same pixels and 3) address what features play the most critical role in delineating these regions.