Estimating Atmospheric Water Vapor Pressure Using Machine Learning
KGE of the Machine Learning model minus that of a standard empirical model which assumes that the dewpoint is given by the minimum air temperature at 83 FLUXNET sites. Red indicates places where the ML model is better; blue places where the empirical model is better.
Bo Gao & Ethan Coon
Atmospheric water vapor pressure (or equivalently, relative humidity) is often not directly measured, but inferred through correlations with other, more frequently measured variables like precipitation and air temperature. Here we develop and evaluate a machine learning method to estimate the vapor pressure, finding that it reduces the failure rate from 32% of sites to 10.9% of sites as compared to more commonly used empirical approaches.
Atmospheric water vapor pressure is an essential forcing variable for simulations of the water cycle. This approach offers great promise for imputing vapor pressure across space and time and could replace older algorithms in frequently used meteorological datasets like DayMet.
This study evaluates the performance of an interpretable Long Short Term Memory (iLSTM) network conditioned on location specific attributes in the estimation of the atmospheric water vapor pressure. This approach allows for training a single transferable model using ensemble data from multiple sites and exploring the quantitative dependency of vapor pressure prediction on multiple environmental variables and their histories. For evaluation, three iLSTM model configurations were setup considering different static variables: no static attribute, mean Leaf Area Index, and mean precipitation. For each configuration, multiple model realizations were trained on meteorologic measurements from 83 FLUXNET sites located in the United States and Canada. Results show that the iLSTM networks significantly improve the prediction in comparison with two empirical methods at most sites, reducing the failure rate from 32% of sites to 10.9% of sites for the best iLSTM model configuration. Additionally, this network provides insight into both the relative importance of the time-series input variables and their temporal importance.