Evaluating functional performance of evapotranspiration models based on causal discovery methods

Investigating causal relationships is critical in ecohydrological systems to understand how variables interact, anticipate future trends, and develop effective models. Traditional predictive performance measures, which compare observations to modeled outputs, cannot fully verify generalization ability and do not address representations of causal linkages between inputs, observations, and outputs. Functional performance focuses on relationships between the target and source variables, but similar to predictive performance metrics, there are multiple ways to quantify these relationships. We compare four causal discovery methods (Granger causality, Transfer Entropy, PCMCI, and Convergent Cross Mapping) to analyze the functional performance of models of evapotranspiration (ET) in an intensively managed agricultural landscape based on eddy covariance measurements over seven years. Evapotranspiration is a critical component of both the energy and water cycles, playing a key role in water resource management and irrigation planning. We first compare methods with linear, nonlinear, and chaotic test cases to evaluate how causal discovery varies with system dynamics, temporal aggregation, and noise. We find that the four methods identify accurate causal relationships to a large extent, but temporal aggregation, and particularly data resampling, can lead to diverging results. Subsequently, we apply the methods to analyze causal sources within three ET models: Priestly-Taylor (PT), Surface Flux Equilibrium (SFE), and Soil Water Balance (SWB) relative to tower observations. The causal discovery methods consistently identify model inputs as primary causal sources, such as net radiation, relative humidity, and temperature for PT and SFE, and soil moisture for SWB. In a functional performance assessment, we rely on the causal strength measures derived from Transfer Entropy and Granger Causality. The functional performance metric offers improved interpretability and flexibility, enabling model evaluation across varying temporal intervals. Among the evaluated models, the functional performance aligns closely with predictive performance, identifying SFE as the best-performing model. We further apply functional performance metrics to evaluate the OpenET dataset, which contains multiple satellite-derived ET products. Although OpenET models rely primarily on satellite-based inputs, they exhibit higher functional performance, indicating a more accurate representation of evapotranspiration processes. The analysis reveals that functional performance metrics offer unique insights into model behavior, aligning with predictive performance, especially when substantial performance differences exist. This study shows how functional performance complements traditional model evaluations, but also highlights sensitivities to different causal inference methods and time scales.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Civil Engineering

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Cao, Jiaze

Contributors dc:contributor

• Kumar, Praveen
• Goodwell, Allison

• Evapotranspiration
• Causal discovery method
• Causal relationship

Statement dc:rights

• Copyright 2025 Jiaze Cao

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129526

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129526