Skip to main content
U.S. flag

An official website of the United States government

A Highly Efficient Deep-Learning Based Parameter Estimation and Uncertainty Reduction Framework for Ecosystem Dynamics Models

Funding Program Area(s)
Project Type
University Grant
Project Term
Project Team

Principal Investigator

Co-Principal Investigator

Project Participant

Our overall goal is to better predict plant responses, such as drought-related changes in evapotranspiration and productivity, by efficiently determining proper plant trait parameters and reducing parametric uncertainties at large scales. We will utilize deep learning (DL) to improve plant trait parameterization for the Energy Exascale Earth System Model (E3SM) Land Model (ELM) coupled to a next-generation dynamic vegetation model, the Functionally Assembled TErrestrial Simulator (FATES).

The project seeks to understand (i) how soil, topography, and other environmental variables influence plant hydraulics, growth, and mortality parameters; (ii) relative importance and uncertainty sources of these parameters; and (iii) characteristics of plant hydraulic parameters, such as time-dependence and identifiability. Building on our previous and preliminary work, a major contribution of the research is a novel and highly efficient parameter learning framework for estimating environment-dependent and scale-dependent vegetation trait parameters with uncertainty estimations. This framework, saving four orders-of-magnitude of computational time, can be applied instantly on large scales, and has very competitive and potentially stronger optimization metrics than current optimization algorithms. It can be said that, instead of solving one inversion problem at a time, our framework will train a tool that solves this problem with higher efficient and more robust results.

Additionally, this framework opens fresh pathways to address issues and ask new questions from a data-driven perspective. We will reduce uncertainty and improve physical significance of parameters by imposing novel constraints, such as information-transfer (parameters are roughly functions of plant types and environmental covariates) and covariation (exploiting observed inter-parameter relationships). At the same time, local heterogeneities in traits can be expressed, allowing us to understand the relative importance of stochasticity and environmental controls of plant traits. By interrogating the trained parameter learning model, we will reveal the linkages between topography, grid scale, soil, climate, and plant hydraulic traits, growth, and mortality parameters. We will also use time-dependent parameters as keys to identify deficiencies in model dynamics. This framework will usher in a new era for parameter determination for land surface models.