Elsevier

Journal of Hydrology

Volume 287, Issues 1–4, 25 February 2004, Pages 214-236
Journal of Hydrology

Using a multiobjective approach to retrieve information on surface properties used in a SVAT model

https://doi.org/10.1016/j.jhydrol.2003.10.003Get rights and content

Abstract

The reliability of model predictions used in meteorology, agronomy or hydrology is partly linked to an adequate representation of the water and energy balances which are described in so-called SVAT (Soil Vegetation Atmosphere Transfer) models. These models require the specification of many surface properties which can generally be obtained from laboratory or field experiments, using time consuming techniques, or can be derived from textural information. The required accuracy of the surface properties depends on the model complexity and their misspecification can affect model performance. At various time and spatial resolutions, remote sensing provides information related to surface parameters in SVAT models or state variables simulated by SVAT models. In this context, the Simple Soil-Plant-Atmosphere Transfer-Remote Sensing (SiSPAT-RS) model was developed for remote sensing data assimilation objectives. This new version of the physically based SiSPAT model simulates the main surface processes (energy fluxes, soil water content profiles, temperatures) and remote sensing data in the visible, infrared and thermal infrared spectral domains. As a preliminary step before data assimilation in the model, the objectives of this study were (1) to apply a multiobjective approach for retrieving quantitative information about the surface properties from different surface measurements and (2) to determine the potential of the SiSPAT-RS model to be applied with ‘little’ a priori information about input parameters. To reach these goals, the ability of the Multiobjective Generalized Sensitivity Analysis (MOGSA) algorithm to determine and quantify the most influential input parameters of the SiSPAT-RS model on several simulated output variables, was investigated. The results revealed the main influential input parameters according to different contrasted environmental conditions, and contributed to the reduction of their a priori uncertainty range. A procedure for specifying surface properties from MOGSA results was tested on the thermal and hydraulic soil parameters, and evaluated through the SiSPAT-RS model performance. Although slightly lower than a reference simulation, the performance were satisfactory and suggested that complex SVAT models can be driven with little a priori information on soil properties, as in a future context of remote sensing data assimilation. Measurement acquired on a winter wheat field of the ReSeDA (Remote Sensing Data Assimilation) experiment were used in this study.

Introduction

Soil Vegetation Atmosphere Transfer (SVAT) models describe energy and water exchanges between the soil, the vegetation and the atmosphere. The understanding and the quantification of these biophysical processes are important for meteorological, agronomical and hydrological purposes. SVAT models require a large set of input parameters and initial state variables that are spatially and temporally distributed. In general, model complexity and the number of parameters increase simultaneously. For example, the Interactions Soil Biosphere Atmosphere (ISBA) model (Noilhan, 1989, Noilhan and J, 1996) used in the operational simulations of the French weather forecast model has around 10 parameters and can easily be applied at different spatial scales. On the other hand, the Simple Soil-Plant-Atmosphere Transfers (SiSPAT) model (Braud et al., 1995) has been developed as a physically based research tool taking into account many biophysical processes, such as detailed profile of root extraction and coupled heat and water exchanges in the soil. This model needs the specification of around 60 parameters and initial state variables (this number depends on the soil description). Most of them vary in time and space and are often assessed through in situ experiments. For example, hydraulic properties in the soil are required to establish the relationship between hydraulic conductivity, soil matric potential, and soil water content. Soil thermal properties are also required. The specification of such soil properties affects significantly model behavior, implying that SVAT models need to be calibrated or regularly corrected.

Generally, calibration in land surface modeling is considered as the determination of a single optimum parameter set, allowing the ‘best’ simulation of several output variables. Due to errors in model structure and parameter uncertainties, the determination of this set is often impossible. Such a difficulty was the basis for the development of multiobjective approaches (Yapo et al., 1998, Gupta, 1998). These approaches are based on statistical analysis of different objective functions in order to determine parameter sets that cannot be distinguished in terms of model performance. The choice and the number of the objective functions depend on the case studied, the model and the available data. For instance, Madsen (2000) used a multiobjective approach to calibrate a conceptual rainfall-runoff model. A single output model variable (discharge) was studied but the author tested and applied different mathematical objective functions in order to simulate different streamflow ranges (peak flow, low flow or overall shape of hydrograph). Gupta et al., 1999, Bastidas et al., 1999 applied the multiobjective approach to the Biosphere Atmosphere Transfer Scheme (BATS, Dickinson, 1984) SVAT model using only one mathematical objective function defined as the Root Mean Square Error (RMSE), but considering simultaneously four different surface variables.

The multiobjective calibration has already been applied in different modeling contexts. Several recent studies have been devoted to the multiobjective calibration of rainfall-runoff models (Madsen, 2000; Boyle et al., 2000, Boyle et al., 2001; Houser et al., 2001) and land surface models (Gupta et al., 1999, Xia et al., 2002, Leplastrier et al., 2002). Such a calibration requires the use of a complex optimization algorithm for retrieving the different parameter sets which lead to a satisfactory simulation of the surface processes. To be efficient, the optimization algorithm generally requires a high number of model runs (Duan et al., 1992, Yapo et al., 1996). In the case of complex models, which have a significant simulation time, this efficiency could be highly impacted.

To reduce the dimension of the parameter optimization problem, multiobjective sensitivity analysis approach has been also investigated in a few recent studies. For example, the Multiobjective Generalized Sensitivity Analysis (MOGSA; Bastidas et al., 1999) algorithm was developed to determine the main influential input parameters. It has already been applied with a land surface model (Bastidas et al., 1999) and a hydrochemical model (Meixner et al., 1999). Although this algorithm was originally developed to determine the model sensitivity to its input parameter, it is also a useful tool for retrieving quantitative information about influential parameters (Demarty, 2001). Such a behavior is potentially interesting for driving complex models which require a significant simulation time.

Moreover, by providing frequent observations at different spatial scales, remote sensing is an attractive tool for retrieving information about surface properties used in land surface models or to control model trajectory. Such methods are referred to as assimilation of remotely sensed data. Even if the assimilation of remotely sensed data has been widely used in atmospheric and oceanographic sciences, it has been implemented only recently in modeling surface processes. This can be done by coupling the SVAT model with Radiative Transfer Models (RTM) in order to simulate both surface processes and remotely sensed data, such as directional reflectances, thermal infrared brightness temperatures, passive microwave brightness temperatures or radar backscattering coefficients. Thus, Burke et al., 1997, Burke et al., 1998 calibrated soil hydraulic properties from L band passive brightness temperature measurements in a soil water and energy budget model, coupled with a microwave emission model (MICROSWEAT). Cayrol et al. (2000) assimilated series of AVHRR surface reflectances and temperatures to calibrate a coupled SVAT-Vegetation growth model. More recently, Wigneron et al. (2002) assimilated L band passive brightness temperature in the ISBA-Ags model (Calvet et al., 1998), coupled with the Tau-Omega microwave radiative transfer model, in order to retrieve initial soil moisture and a parameter controlling vegetation growth.

This article presents a multiobjective approach which was performed on the Simple Soil-Plant-Atmosphere Transfer-Remote Sensing (SiSPAT-RS) model. This new version of the SiSPAT model (Braud et al., 1995) was developed to facilitate the assimilation of remotely sensed data. It simulates soil-vegetation-atmosphere energy and water transfers as well as remote sensing measurements, at the field scale, in the visible–near infrared and thermal infrared spectral domains. As a preliminary step before data assimilation, the objectives of this study were (1) to apply a multiobjective approach for retrieving substantial information about the surface properties from surface measurements and (2) to determine the potential of the SiSPAT-RS model to be applied with ‘little’ a priori information about input parameters. To reach these goals, this article focuses on the potential of the MOGSA algorithm to determine and quantify the most influential input parameters of the SiSPAT-RS model. Its implementation was tested on a winter wheat field, from the data base collected during the Alpilles-ReSeDA (Remote Sensing Data Assimilation) experiment.

This article is divided into six sections. Section 2 provides an overview of the Alpilles-ReSeDA experiment and the particular data set used in this study. Section 3 describes the SiSPAT-RS model. Section 4 presents the multiobjective approach which was conducted with the MOGSA algorithm and its implementation. Section 5 describes the results. Finally, Section 6 opens the article for discussion and presents some conclusions.

Section snippets

The Alpilles-ReSeDA experiment

The aim of the Alpilles-ReSeDA experiment (Olioso et al., 2002a) was to provide a consistent dataset for assessing crop and soil processes using remote sensing data. This experiment was focused on agricultural land and practices in order to develop ReSeDA methods to better evaluate soil and vegetation processes (biomass production, crop yield, energy and water budgets). Therefore, a small agricultural area, characterized by a large diversity of crops, was instrumented and monitored during one

Modeling approach

In this study, the SiSPAT model (Braud et al., 1995, Braud, 2000) with a Remote Sensing (SiSPAT-RS) module (Demarty, 2001, Demarty et al., 2002) was used. It was developed specifically to assess the value of multi-spectral remotely sensed data within a modeling framework. The ultimate objective of this work is to improve the prediction of the SVAT prognostic variables by assimilation of remote sensing data. For this purpose, the SVAT model was coupled with two RTM, in the visible–near infrared

Multiobjective methodology

A multiobjective framework was established in order to progress toward the objectives of retrieving information about the surface properties used in the SiSPAT-RS model. The framework combined a sensitivity analysis and a parameter retrieval procedure. Unfortunately, the use of a complex optimization algorithm was not consistent with the large computer time required by the SiSPAT-RS model. An alternative procedure involved evaluation of the ability of the MOGSA algorithm (Bastidas et al., 1999)

Multiobjective sensitivity analysis

Samples of maximum 2500 simulations were realized for the three studied periods. Tests were performed to verify that the size of the samples was large enough to obtain robust results. They showed that a number of 1500 simulations was the minimum sample size necessary both to stabilize the number of sensitive parameters and to eliminate the impact of initial parameter sampling on the sensitivity analysis results. Furthermore, a Pareto rank 2 was chosen as threshold for the first specific period,

Conclusion

This study presents a multiobjective framework which was applied to the SiSPAT-RS. This model was developed to simulate the main surface processes (energy fluxes, soil water content, temperatures) and remote sensing observations in the visible–infrared and thermal infrared spectral domains. For the visible–infrared wavebands, the Multi-layer and Multi-element (2M) version (Weiss et al., 2001) of the Scattering by Arbitrary Inclined Leaves (SAIL; Verhoef, 1984, Verhoef, 1985) canopy radiative

Acknowledgements

The authors would like specially to thank all the partners of the ReSeDA program for providing ground truth data. This work was supported by the INSU/CNRS French national programs (PNRH and PNTS). We are pleased to thank Doctor Pierre Guillevic and anonymous reviewers for very constructive suggestions.

References (47)

  • M Weiss et al.

    Coupling canopy functioning and radiative transfer models for remote sensing data assimilation

    Agricultural and Forest Meteorology

    (2001)
  • P.O Yapo et al.

    Automatic calibration of conceptual rainfall-runoff models: sensitivity to calibration data

    Journal of Hydrology

    (1996)
  • P.O Yapo et al.

    Multi-objective global optimization for hydrologic models

    Journal of Hydrology

    (1998)
  • L.A Bastidas et al.

    Sensitivity analysis of a land surface scheme using multicriteria methods

    Journal of Geophysical Research

    (1999)
  • D.P Boyle et al.

    Toward improved calibration of hydrologic models: combining the strengths of manual and automatic methods

    Water Resources Research

    (2000)
  • D.P Boyle et al.

    Toward improved streamflow forecasts: value of semidistributed modeling

    Water Resources Research

    (2001)
  • Braud I. SiSPAT, a numerical model of water and energy fluxes in the soil-plant-atmosphere continuum. Version 3.0,...
  • I Braud et al.

    Soil properties, initial and boundary conditions for use within SVAT models in the framework of the intercomparison of SVAT models used in the Alpilles-ReSeDA project, Alpilles-ReSeDA data base

    (2000)
  • R.H Brooks et al.

    Corey, Hydraulic Properties of Porous Media

    (1964)
  • E.J Burke et al.

    Calibrating a soil water and energy budget model with remotely sensed data to obtain quantitative information about the soil

    Water Resources Research

    (1997)
  • E.J Burke et al.

    Using a modeling approach to predict soil hydraulic properties from passive microwave measurements

    IEEE Transactions on Geoscience and Remote Sensing

    (1998)
  • P Cayrol et al.

    Calibrating a coupled SVAT-Vegetation growth model with remotely sensed reflectance and surface temperature—a cas study for the HAPEX-Sahel grassland sites

    Journal of Applied Meteorology

    (2000)
  • J.W Deardorff

    Efficient prediction of ground surface temperature and moisture, with inclusion of a layer of vegetation

    Journal Geophysics Research

    (1978)
  • Cited by (51)

    • A combination of DISPATCH downscaling algorithm with CLASS land surface scheme for soil moisture estimation at fine scale during cloudy days

      2016, Remote Sensing of Environment
      Citation Excerpt :

      Land surface schemes have been designed to simulate the exchange of energy and water between soils, vegetation, and the atmospheric layer. They were always considered as a good alternative to estimate soil moisture with the benefit of being operational at different spatial and temporal scales regardless of soil and weather conditions (Demarty et al., 2004). Version 3.5 of the Canadian Land Surface Scheme (CLASS; Verseghy, 2011) was used since it is assumed to be better than others for the Canadian prairies (Henderson & Yang, 1993; Comer et al., 2000; Alavi et al., 2010).

    • A rule-based image analysis approach for calculating residues and vegetation cover under field conditions

      2014, Catena
      Citation Excerpt :

      A major step in providing more objectivity in the selection process of fitting rules would be a procedure of automated fitting of multiobjective criteria, as already implemented for a wide range of environmental modelling applications (Demarty et al., 2004; Tang et al., 2007; Vrugt et al., 2003).

    • Land surface temperature product validation using NOAA's surface climate observation networks-scaling methodology for the visible infrared imager radiometer suite (VIIRS)

      2012, Remote Sensing of Environment
      Citation Excerpt :

      Model calibration consists of the minimization of a cost function expressing the divergence between model outputs and observations. In order to achieve a robust model calibration, the multi-objective calibration iterative process (MCIP; Demarty et al., 2004, 2005) was applied. Based on a stochastic Monte Carlo approach, the MCIP methodology represents the reduction of an initial parameter space (initial parameter ranges are defined in Table 2) by the optimization of one or several model outputs against observations, e.g., LST, surface fluxes, etc.

    • Factors controlling the isotopic partitioning between soil evaporation and plant transpiration: Assessment using a multi-objective calibration of SiSPAT-Isotope under controlled conditions

      2012, Journal of Hydrology
      Citation Excerpt :

      Calibrated state-of-the-art SVAT models such as SiSPAT-Isotope in this context can inform on soil–plant and -atmosphere interactions. Multi-objective calibration tools have been substantially used by the hydrological community for nearly a decade (e.g. Madsen, 2003; see review of Efstratiadis and Koutsoyiannis, 2010) and to a lesser extent by the SVAT modelling community (e.g. Bastidas et al., 1999; Coudert and Ottlé, 2007; Demarty et al., 2004, 2005) but, to our knowledge, never for laboratory experiments. From the multi-objective calibration results, we could emphasize the critical need of further knowledge on isotopic parameters controlling the isotopic partitioning of ET such as the kinetic fractionation factor, used in the determination of the isotopic composition of soil evaporation in the field for evapotranspiration partitioning studies (Moreira et al., 1997; Yepez et al., 2003, 2005; Wang et al., 2010; Wang and Yakir, 2000).

    View all citing articles on Scopus
    View full text