Marco probabilístico para la estimación de evapotranspiración en huertos frutales: integración de modelos físicos y estructuras de error espaciotemporales
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La estimación de la evapotranspiración (ET) es fundamental para la gestión del agua en sistemas agrícolas bajo presión hídrica creciente. Las plataformas satelitales disponibles operativamente presentan sesgos sistemáticos en huertos frutales y reportan exclusivamente estimaciones puntuales sin cuantificación de incertidumbre, lo que limita la capacidad de usuarios y gestores para evaluar la confiabilidad de los productos disponibles. Para abordar esta brecha, se desarrollo un marco metodológico probabilístico en dos fases orientadas a construir un ensamble con diversidad estructural. El modelo SEB-PV (desarrollado exclusivamente para superficies parcialmente vegetadas) fue identificado como candidato complementario a los seis algoritmos de la plataforma OpenET, pero requería superar barreras operacionales. La primera fase evaluó su viabilidad utilizando productos globales de humedad del suelo y datos de estaciones agrometeorológicas, comparando siete algoritmos de escalamiento temporal. La segunda fase integro el modelo operacionalizado con OpenET en un marco de Promedio de Modelos Bayesiano, contrastando arquitecturas de calibración bajo independencia condicional versus estructuras de error espaciotemporales explícitas, evaluadas mediante validación cruzada temporal con encadenamiento hacia adelante.
Los resultados de la primera fase indicaron que SEB-PV puede operar con precisión aceptable utilizando datos globalmente disponibles, aunque no se identifico un algoritmo de escalamiento temporal universal, siendo su desempeño dependiente del contexto climático y estructural del huerto. En la segunda fase, la especificación explícita de la estructura de error espaciotemporal resulto el factor dominante de la exactitud predictiva fuera de muestra, superando consistentemente a los marcos de independencia condicional en todas las arquitecturas evaluadas. La calibración de intercepto dependiente del estado resulto la arquitectura con mayor exactitud predictiva, con una mejora relativa del 46% en error normalizado respecto al promedio determinístico simple, superando el umbral planteado en la hipótesis, y mostró mayor estabilidad frente a condiciones meteorológicas no observadas durante la calibración. Esta investigación establece que la especificación de la estructura de error es el determinante principal de la exactitud predictiva en ensambles satelitales de ET, orientando donde concentrar esfuerzos metodológicos. El marco probabilístico desarrollado aborda ambas limitaciones identificadas: reduce el sesgo sistemático mediante calibración adaptativa y produce distribuciones predictivas posteriores que cuantifican la incertidumbre asociada a cada estimación, proveyendo a usuarios y gestores una herramienta para discernir cuando confiar y cuándo cuestionar los productos satelitales. El código Stan está disponible públicamente para facilitar su adopción y adaptación a otros sistemas.
Evapotranspiration (ET) estimation is fundamental for water management in agricultural systems under increasing hydrological pressure. Operationally available satellite-based ET platforms exhibit systematic biases in fruit orchards and report exclusively point estimates without uncertainty quantification, limiting the ability of users and water managers to evaluate the reliability of available products. To address this gap, a probabilistic methodological framework was developed in two phases oriented toward building an ensemble with structural diversity. The SEB-PV model (developed exclusively for partially vegetated surfaces) was identified as a complementary candidate to the six algorithms of the OpenET platform, but required overcoming operational barriers. The first phase evaluated its viability using global gridded soil moisture products and agrometeorological station data, comparing seven temporal upscaling algorithms. The second phase integrated the operationalized model with OpenET within a Bayesian Model Averaging framework, contrasting calibration architectures under conditional independence versus explicit spatiotemporal error structures, evaluated through forward-chaining temporal crossvalidation. Results from the first phase indicated that SEB-PV can operate with acceptable accuracy using globally available data, although no universal temporal upscaling algorithm was identified, with performance being dependent on the climatic and structural context of the orchard. In the second phase, explicit specification of the spatiotemporal error structure emerged as the dominant driver of out-of-sample predictive accuracy, consistently outperforming conditional independence frameworks across all evaluated architectures. State-dependent intercept calibration yielded the highest out-of-sample predictive accuracy, with a 46% relative improvement in normalized error over the deterministic simple average, surpassing the threshold stated in the hypothesis, and showed greater stability under weather conditions not observed during calibration. This research establishes that error structure specification is the primary determinant of predictive accuracy in satellitebased ET ensembles, guiding where methodological efforts should be concentrated. The probabilistic framework developed addresses both identified limitations: it reduces systematic bias through adaptive calibration and produces posterior predictive distributions that quantify the uncertainty associated with each estimate, providing users and water managers a tool to discern when to trust and when to question satellite products. The Stan code is publicly available to facilitate adoption and adaptation to other systems.
Evapotranspiration (ET) estimation is fundamental for water management in agricultural systems under increasing hydrological pressure. Operationally available satellite-based ET platforms exhibit systematic biases in fruit orchards and report exclusively point estimates without uncertainty quantification, limiting the ability of users and water managers to evaluate the reliability of available products. To address this gap, a probabilistic methodological framework was developed in two phases oriented toward building an ensemble with structural diversity. The SEB-PV model (developed exclusively for partially vegetated surfaces) was identified as a complementary candidate to the six algorithms of the OpenET platform, but required overcoming operational barriers. The first phase evaluated its viability using global gridded soil moisture products and agrometeorological station data, comparing seven temporal upscaling algorithms. The second phase integrated the operationalized model with OpenET within a Bayesian Model Averaging framework, contrasting calibration architectures under conditional independence versus explicit spatiotemporal error structures, evaluated through forward-chaining temporal crossvalidation. Results from the first phase indicated that SEB-PV can operate with acceptable accuracy using globally available data, although no universal temporal upscaling algorithm was identified, with performance being dependent on the climatic and structural context of the orchard. In the second phase, explicit specification of the spatiotemporal error structure emerged as the dominant driver of out-of-sample predictive accuracy, consistently outperforming conditional independence frameworks across all evaluated architectures. State-dependent intercept calibration yielded the highest out-of-sample predictive accuracy, with a 46% relative improvement in normalized error over the deterministic simple average, surpassing the threshold stated in the hypothesis, and showed greater stability under weather conditions not observed during calibration. This research establishes that error structure specification is the primary determinant of predictive accuracy in satellitebased ET ensembles, guiding where methodological efforts should be concentrated. The probabilistic framework developed addresses both identified limitations: it reduces systematic bias through adaptive calibration and produces posterior predictive distributions that quantify the uncertainty associated with each estimate, providing users and water managers a tool to discern when to trust and when to question satellite products. The Stan code is publicly available to facilitate adoption and adaptation to other systems.
Description
Tesis presentada para optar al grado de Doctor en Recursos Hídricos y Energía para la Agricultura
Keywords
Evapotranspiración, Agricultura de precisión -- Chile, Riego -- Eficiencia, Huertos frutales