This site requires Cookies enabled in your browser for login.
. . .
Alias List Editor
AOSM2022: Posterior-informed feature importance method for examining how large-scale climatic indices influence hydrological processes in Continental US
Section 1: Publication
Authorship or Presenters
kailong Li, Saman Razavi
Posterior-informed feature importance method for examining how large-scale climatic indices influence hydrological processes in Continental US
Hydrology and Terrestrial Ecosystems
kailong Li, Saman Razavi (2022). Posterior-informed feature importance method for examining how large-scale climatic indices influence hydrological processes in Continental US. Proceedings of the GWF Annual Open Science Meeting, May 16-18, 2022.
AOSM2022 Next Generation Hydrological Modelling
Section 2: Abstract
Plain Language Summary
Feature importance has been widely used for machine learning models to examine the relative significance of model predictors. This study developed a posterior-informed feature importance method (PIFI) for hydrological inference. The proposed PIFI is based on the bootstrap aggregated Wilks statistics and stratified sampling Bayesian model averaging (SSBMA). Each Wilks statistics is considered an ensemble member of SSBMA, and its posterior probabilities are evaluated based on a spectrum of streamflow quantile ranges. Compared with conventional feature importance methods such as permutation feature importance (PFI) and mean decrease in impurity (MDI), the proposed PIFI can help investigate the varying significance of a predictor in response to the variations of streamflow. PIFI has been applied to the Catchment Attributes and Meteorological (CAMELS) dataset, which contains forcing and hydrologic response data for 673 basins across the contiguous United States that spans a very wide range of hydroclimatic conditions. In attempting to demonstrate the relative importance of meteorological data and large-scale climatic indices on streamflow, we used monthly mean values of meteorological data in CAMELS dataset and four commonly used large-scale climatic indices (including Nino3.4, Pacific decadal oscillation (PDO), interdecadal Pacific oscillation (IPO) and Pacific North American index (PNA)) to simulate the monthly streamflows. Our results suggest that Nino3.4 strongly influences both low and high flows, whereas IPO indicates the most substantial influence over median flows. These results can help us better understand how drought and flood may be associated with large-scale climatic indices.
Section 3: Miscellany
Global institute for water security
First Author: kailong Li, Global institute for water security
Additional Authors: Saman Razavi, Global institute for water security
Section 4: Download
T-2022-04-24-11sibSfNSuUyhsInmc5auIQ Conference Publication 1.0