H191-05
Hypothesis Testing using Long Short-Term Memory Networks Applied to Large Pixel-Scale Datasets
Tuesday, 15 December 2020: 20:46
Virtual
Yuan-Heng Wang1, Hoshin Gupta2, Grey Stephen Nearing3, Xubin Zeng1 and Guo-Yue Niu4, (1)University of Arizona, Department of Hydrology and Atmospheric Sciences, Tucson, AZ, United States, (2)Hydrology and Atmospheric Sciences, The University of Arizona, Tucson, AZ, United States, (3)Natel Energy Inc, Upstream Tech, Alameda, CA, United States, (4)University of Arizona, Hydrology and Atmospheric Sciences, Tucson, AZ, United States
Abstract:
Recent studies have shown the Long Short-Term Memory (LSTM) network to be capable of extrapolating knowledge to ungauged spatial locations. Here, we use LSTM to investigate different structural hypotheses 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 the network inputs are progressively augmented, mainly as suggested by regional random forest regression models, with various levels of spatial (regional and local) information to improve predictive performance and parameter identifiability. These levels include, 1) regional prior information in the form of physiographical region ID or regional categorization, 2) prior information plus local (pixel-scale) ancillary static (land cover type, vegetation, physiographic, cloud and snow properties) features, and 3) prior information, local static features and local dynamic variables.
We show the error (Nash Sutcliffe Efficiency and total bias) vs. epoch plot as well as SWE prediction bias over CONUS. The experiments investigate the extent to which different spatial regions can share a common dynamical model structure with different parameter values. We are particularly interested in how LSTMs can be used to test system structure hypotheses, imposed in the form of different kinds of regularization strategies, thereby providing insights into the issue of uniqueness of place.
