Optimization of a remote sensing energy balance method over different canopy applied at global scale

https://doi.org/10.1016/j.agrformet.2019.107633Get rights and content

Highlights

  • Key parameters specific to different biomes in SEBS model were identified.

  • Key parameters for seven biomes were optimized by using particle swarm optimization.

  • Optimized SEBS model (SEBS-opt) produces an accurate global daily ET estimate.

Abstract

Parameterization methods which calculate turbulent heat and water fluxes with thermal remote sensing data were evaluated in the revised remote sensing surface energy balance system (SEBS) model (Chen et al., 2013). The model calculates sensible heat (H) based on the Monin-Obukhov similarity theory (MOST) and determines latent heat (LE) as the residual of energy balance. We examined the uncertainties of H and LE in the SEBS model due to five key parameters at the local station point scale. Observations at 27 flux towers located in seven land cover types (needle-leaf forest, broadleaf forest, shrub, savanna, grassland, cropland, and sparsely vegetated land) and an artificial intelligence particle swarm optimization (PSO) algorithm was combined to calibrate the five parameters (leaf drag coefficient, leaf heat transfer coefficients, roughness length for soil, and two parameters for ground heat calculation) in the SEBS model. The root-mean-square error at the site scale was reduced by 9 Wm−2 for H, and 92 Wm−2 for LE, and their correlation coefficients were increased by 0.07 (H) and 0.11 (LE) after using the calibrated parameters. The updated model validation was further conducted globally for the remotely sensed evapotranspiration (ET) calculations. Overestimation of SEBS global ET was significantly improved by using the optimized values of the parameters. The results suggested PSO was able to consistently locate the global optimum of the SEBS model, and appears to be capable of solving the ET model optimization problem.

Introduction

The remote sensing surface energy balance (SEBS) model predicts the magnitude of land surface fluxes and evapotranspiration (ET) by simulation of canopy-atmosphere turbulent physical processes. The SEBS model performance has not always been acceptable at the evaluation flux sites (Chen et al., 2019, 2013; Ershadi et al., 2014; Ma et al., 2018; Michel et al., 2016). SEBS simulation of ET at the single-point flux stations have been intensively done by WACMOS-ET project (Michel et al., 2016), meanwhile no clear reasons for overestimation of ET are given. The SEBS model showed low performance over tall and heterogeneous canopies, which was likely a consequence of the effects of the roughness sub-layer parameterization is not employed in the model. Chen et al. (2019) reported that sensible heat is significantly underestimated by SEBS at forest sites due to a high value of excess resistance (kB-1) and introduced vertical foliage density information into a column canopy-air turbulent diffusion model, verified that the new model can provide an accurate simulation over different canopies. There is no agreement in previous research regarding the main cause of the model uncertainties. These uncertainties have either been attributed to: (i) errors in the roughness parameters (Timmermans et al., 2013), (ii) either the land surface temperature (Kwast et al., 2009) or the land-air temperature gradient, (iii) the vegetation fraction, and/or (iv) the wet limit criteria (Gibson et al., 2011), (v) vulnerabilities associated with the partitioning of the available energy (Webster et al., 2017a). Ma et al. (2018) did a sensitivity analysis for SEBS by using a random-sampling high-dimensional model representation approach. Their results indicated that net radiation (Rn), ground heat flux (G0), heat and momentum roughness (z0m,z0h), and land surface temperature (LST) had the most impact on surface energy fluxes and ET estimation in SEBS model. The uncertainness in Rn, G0 and LST are associated to that of satellite and meteorological input data, and not related to the model itself. There are also some studies on improvements of the model and the model parameters (e.g. z0m,z0h and excess resistance kB-1). Webster et al. (2017b) introduced energy restraint process into the SEBS and improved the model accuracy at a heavily forested site and a sub-alpine grassland site. Gokmen et al. (2012) integrate water stress into the calculation of sensible heat, through a soil moisture modified kB-1, to overcome the underestimation of sensible heat flux. Chen et al. (2013) advanced SEBS kB-1 to make the model has an ability to accurately estimate sensible heat flux for bare soil, sparse canopy, dense canopy, and snow surface. Previous studies suggested modifications to the critical parameters in the excess resistance (kB-1) calculation (Chen et al., 2019, 2013; Gokmen et al., 2012; Timmermans et al., 2013), which in turn provided more realistic simulations for the turbulent flux in semi-arid areas (Chen et al., 2013; Hong et al., 2012) and agricultural land (Yang et al., 2010). In order to produce a global daily ET by SEBS, Vinukollu et al. (2011) used a different heat/momentum roughness length parameterization method from the SEBS model. The accuracy of heat flux predictions depend on how the model parameters are determined. Meanwhile, the model parameters are not directly measurable and must be estimated through model calibration, i.e. by fitting the model outputs to their observations. Consequently, there is a clear need to have an effective and efficient optimization procedure that can help in the automatic identification of a realistic set of optimal parameters for the model.

It is found that canopy parameters (Cd, Ct, C1) and the two empirically derived parameters (Γs and Γc) in SEBS are the most important parameters for H and LE calculation. Cd, Ct, and C1 have physically meaningful interpretations, but possess some uncertainties. Massman (1997) suggested the possibility to alter these parameters to get a better flux simulation by drawing on a large observation dataset for vegetated surfaces. In order to describe the impact of canopy geometry on roughness length and turbulent heat flux, the values for the parameters need to be verified over different canopies. When land surface changes from bare soil, to grass, shrub and forest canopy, the frictional resistance to surface airflow will be increased, so the roughness of the underlying surface also increases. The roughness parameters estimated by LAI in SEBS may not be sufficient for different vegetated land surfaces. Site to site variations in canopy roughness may be represented by an effective parameter value. Based on the sensitivity evaluations, we suggest that different parameter values should be used for different sites according to their canopy classifications.

Determination of model parameters requires simultaneously solving a set of non-linear equations. The calculation is complex, iterative and time consuming. Consequently, scientists have been exploring ways to include 'expert knowledge' into the automatic parameter calibration procedures. A local search procedure has very low probability of successfully finding the optimal parameter sets (Duan et al., 1992). Therefore, researchers have explored the use of global optimization methods for model calibration (Cooper et al., 1997; Duan et al., 1992). The recent development of the particle swarm optimization (PSO) algorithm used in artificial intelligence provides one possible method for global optimization (Kennedy and Eberhart, 1995). The PSO algorithm mimics animal activities such as the processes birds and fish use for finding food. Particles constrained by a certain object function in PSO enable a global optimal solution within a given space. Many studies have used this method to calibrate the parameters of continental hydrological models. Yang et al. (2017) used the PSO algorithm to calibrate the empirical parameters in the Monin-Obukhov similarity theory (MOST). Gill et al. (2006) used a multi-objective particle swarm algorithm to estimate hydrological parameters. Chau (2006) combined the PSO algorithm with artificial neural networks (ANNs) to predict water levels. Scheerlinck et al. (2009) used the PSO algorithm to calibrate the parameters of a simple hydrological model. These results demonstrate that the PSO scheme yields reasonable parameter values for their models. Calibrating the turbulent flux parameters in the SEBS model was accomplished with a similar optimization process. We compared the errors between the calculated and measured values of the sensible/latent heat fluxes within a time period longer than one year, evaluated the suitability of the parameters, and then obtained more accurate parameters. With a view to improving the model performance, we demonstrate and recommend some modifications of the critical parameters that achieved significant improvements in the H and LE simulations.

The combination of PSO and SEBS in this study has a broad application to other ET models, e.g. Penman-Monteith (Cleugh et al., 2007; Monteith, 1965; Mu et al., 2007; Zhang et al., 2010), Priestley-Taylor (Fisher et al., 2008; Jiang and Islam, 2001; Miralles et al., 2011a; Priestley and Taylor, 1972), and LSA-SAF ET models (Ghilain et al., 2011). Observations from the flux stations will also play an important role for PSO to calibrate the ET model parameters. Section 2 introduce the methodology on how to combine PSO and SEBS model to do parameter calibration, and flux dataset used in this study. SEBS model and PSO algorithm were also introduced in Section 2. We first do the model simulation in Section 3, by using meteorological measurements at twenty-seven flux tower which are representative of seven land cover units. The simulated fluxes were compared to the flux tower measurements. A sensitivity analysis was also performed in Section 3 to identify the key parameters that were responsible for most of the variability in the heat/water flux results. Observations from the twenty-seven flux stations were employed by the PSO to calibrate the key parameters. The optimized SEBS (SEBS-opt) model was then used to estimate global daily ET by incorporating MODIS satellite data, along with an evaluation of the remote sensing ET product. Section 4 discusses these results and makes conclusions.

Section snippets

General approach

We evaluated the impact of varying critical model parameters on both H and LE flux estimates, and used this information to identify the causes for biases in H and LE estimations. The values for the SEBS model parameters, (e.g. C1, Cd, Ct, Γs and Γc described in Section 2.4), cannot be measured directly. We use H and LE model output and their measurement to do the model parameter sensitivity analysis. The objective of the parameter optimization was to produce lower root‐mean‐square error (RMSE),

Documenting problems of the model over different canopies

In this section, we analyzed the model structure related to errors in H and LE simulations to provide a baseline for comparison with optimized results. Fig. 2 shows H and LE comparisons between flux tower measurements and model simulations using the SEBS at seven sites. The seven sites are selected out from the seven land cover types respectively, due to the figure can not hold the results for the 27 sites. Note that good correspondence of measured and modeled values should place values along

Discussion

It is difficult to determine five parameters simultaneously in the SEBS model by solving a set of nonlinear equations, which is the approach previously employed. The artificial intelligence algorithm provides a feasible approach to solving this problem. With measurements from Fluxnet stations as a basis, this study used the PSO algorithm to calibrate the model parameters associated with diverse land surface types. This study concluded that by using the PSO algorithm, the model parameters can be

Conclusions

The SEBS model has been validated and calibrated to be suitable for both tall and short canopies. This model was driven by a time series of meteorological observation data at point scales and its performance has been evaluated by comparisons between its simulated and observed fluxes. The measurements were performed over seven typical land-cover types that cover the major land areas globally. The results show that the roughness parameterization method previously employed tends to produce

Acknowledgments

Xuelong Chen was supported by CAS Pioneer Hundred Talents Program. We acknowledge the following AmeriFlux sites: site Us-NC2, US-GLE, US-MRf, US-NR1, US-UMB, US-ChR, US-MOz, US-MMS, US-WBW, US-WCr, US-Aud, US-Me6, US-Fmf, Us-Br1, US-Ctn, US-Wkg, the following TERN OzFlux sites: Gingin, Calperum, Ti Tree East, Riggs Creek, Daly Pasture, and Arcturus, for providing us with their data sets. AmeriFlux, OzFlux, European Fluxes Database Cluster, ChinaFlux are acknowledged for the in-situ data. The

References (56)

  • Q. Mu et al.

    Development of a global evapotranspiration algorithm based on MODIS and global meteorology data

    Remote Sens. Environ.

    (2007)
  • R.H. Shaw et al.

    Aerodynamic roughness of a plant canopy: a numerical experiment

    Agr. Meteorol.

    (1982)
  • R.K. Vinukollu et al.

    Global estimates of evapotranspiration for climate studies using multi-sensor remote sensing data: evaluation of three process-based approaches

    Remote Sens. Environ.

    (2011)
  • E. Webster et al.

    Incorporating an iterative energy restraint for the Surface Energy Balance System (SEBS)

    Remote Sens. Environ.

    (2017)
  • E. Webster et al.

    Incorporating an iterative energy restraint for the Surface Energy Balance System (SEBS)

    Remote Sens. Environ.

    (2017)
  • Q. Yang

    Using the particle swarm optimization algorithm to calibrate the parameters relating to the turbulent flux in the surface layer in the source region of the Yellow River

    Agric. For. Meteorol.

    (2017)
  • P.O. Yapo et al.

    Multi-objective global optimization for hydrologic models

    J. Hydrol.

    (1998)
  • M.A. Abido

    Optimal design of power-system stabilizers using particle swarm optimization

    IEEE Trans. Energy Convers.

    (2002)
  • J. Arnqvist et al.

    Flux-profile relation with roughness sublayer correction

    Q. J. R. Meteorol. Soc.

    (2015)
  • F.C. Bosveld

    Exchange Processes Between a Coniferous Forest and the Atmosphere

    (1999)
  • G.G. Burba et al.

    Novel design of an enclosed CO2/H2O gas analyser for eddy covariance flux measurements

    Tellus B

    (2010)
  • X. Chen et al.

    A column canopy-air turbulent diffusion method for different canopy structures

    J. Geophys. Res. Atmos.

    (2019)
  • X. Chen

    Development of a 10-year (2001–2010) 0.1° data set of land-surface energy balance for mainland China

    Atmos. Chem. Phys.

    (2014)
  • X. Chen

    An improvement of roughness height parameterization of the Surface Energy Balance System (SEBS) over the Tibetan Plateau

    J. Appl. Meteorol. Climatol.

    (2013)
  • Q. Duan et al.

    Effective and efficient global optimization for conceptual rainfall-runoff models

    Water Resour. Res.

    (1992)
  • J. Fisher et al.

    Global evapotranspiration: a critical variable linking ecosystem functioning, carbon and climate feedbacks, agricultural management, and water resources

    AGU Water Resourced Res.

    (2016)
  • J.B. Fisher

    The future of evapotranspiration: global requirements for ecosystem functioning, carbon and climate feedbacks, agricultural management, and water resources

    Water Resou. Res.

    (2017)
  • P. Gentine et al.

    The diurnal behavior of evaporative fraction in the soil–vegetation–atmospheric boundary layer continuum

    J. Hydrometeorol.

    (2011)
  • Cited by (22)

    • Evapotranspiration estimation using a modified crop coefficient model in a rotated rice-winter wheat system

      2022, Agricultural Water Management
      Citation Excerpt :

      Many methods of ETc measurement and estimation have been developed (Allen et al., 2011b; Paredes et al., 2018). To date, the ETc can be directly or indirectly measured from a wide range of measurement systems including soil water balance (Qiu et al., 2015; Rashid Niaghi and Jia, 2019; Pereira et al., 2020a; Chen et al., 2021), lysimeters (Ding et al., 2010; Kim et al., 2011; Duncan et al., 2016; López-Urrea et al., 2020), sap flow plus micro-lysimeters (Zhang et al., 2011b; Rafi et al., 2019), Bowen-ratio energy balance (Qiu et al., 2019; Rashid Niaghi and Jia, 2019), eddy covariance (Zhang et al., 2016; Liu et al., 2019; Rashid Niaghi et al., 2019; Olivera-Guerra et al., 2020; Er-Raki et al., 2021), scintillometry (optical scintillometer plus /or microwave scintillometer) (Ezzahar et al., 2009; Ezzahar and Chehbouni, 2009; Vendrame et al., 2020), remote sensing energy balance (Song et al., 2016; Chen et al., 2019), and satellite-based ETc using vegetation indices methods (Allen et al., 2011b). Among these methods, the Bowen-ratio energy balance method is a relatively reliable and practical technology for measuring ETc (Allen et al., 2011b), and has been widely used in a variety of field conditions (Inman-Bamber and McGlinchey, 2003; Zhang et al., 2008; Rashid Niaghi and Jia, 2019).

    • An improved remote sensing based approach for predicting actual Evapotranspiration by integrating LiDAR

      2021, Advances in Space Research
      Citation Excerpt :

      Specifically, the SMC site being in temperate zone showed relatively large overestimation for LE (111.15 Wm−2), and relatively larger underestimation for H estimations (−151.91 Wm−2). The underestimation of H and overestimation of LE in the forest ecosystem have been reported in previous studies (Chen et al., 2019; Michel et al., 2016; Liaqat and Choi, 2015; Ershadi et al., 2014; Byun et al., 2014), and are attributed to overestimation of the excess resistance in the fractions of canopies used in the SEBSLandsat. Such overestimation of excess resistance may be related to the current parameterization of hc, which directly influences the roughness length parameter, including z0m.

    View all citing articles on Scopus
    View full text