A Robust Light Use Efficiency Model Parameterization Method Based on
Ecosystem Properties
Abstract
In a model simulating the dynamics of a system, parameters can represent
system properties and unresolved processes, therefore affecting the
model accuracy and uncertainties. For a light use efficiency (LUE)
model, which is a typical tool to estimate gross primary productivity
(GPP), the plant-functional-type (PFT)-dependent parameterization method
was widely used to extrapolate parameters to larger spatial scales.
However, the method cannot capture the spatial variability within PFT
and introduces misclassification errors. To overcome the shortage, here
we proposed an ecosystem-property-based parameterization method
(mNN-GPP) for an LUE model. This method refers to predicting model
parameters using the multi-output artificial neural network based on
collected variables including PFT, climate types, bioclimatic variables,
vegetation features, atmospheric deposition and soil properties at 196
FLLUXNET eddy covariance flux sites. The neural network was optimized
according to GPP errors and constraints on sensitivity functions of the
LUE model. We compared mNN-GPP with eleven other typical parameter
extrapolating methods, including PFT-, climate-specific
parameterization, global and PFT-based parameter optimization,
site-similarity-based, and regression methods. These twelve methods were
assessed using Nash-Sutcliffe model efficiency (NSE), determination
coefficient and normalized root mean squared error of the simulated GPP.
The simulated results were also contrasted with those of site-specific
calibration based on full-time-series GPP estimated from observational
net ecosystem exchange. The N-fold cross-validated results showed that
mNN-GPP had the best performance across various temporal and spatial
scales (e.g., NSE=0.62 at the daily scale). No extrapolated parameters
reached the same performance as the calibrated parameters (NSE=0.82),
but the ranges of predicted parameters were constrained. Furthermore,
the Shapley values, layer-wise relevance and partial dependence of the
input features showed that bioclimatic variables, PFT, and vegetation
features are the key variables determining parameters. We recommend
using the parameterization method considering both ecosystem properties
and prediction errors to other GPP models and across spatio-temporal
scales.