Estimating the dynamics of ecosystem CO2 flux and biomass production in agricultural fields on the basis of synergy between process models and remotely sensed signatures



[1] The objective of this study was to investigate the potential of synergy between biophysical/ecophysiological models and remote sensing for the dynamic estimation of biomass and net ecosystem exchange of CO2 (NEECO2). We obtained a long-term data set of micrometeorological, plant, and remote sensing measurements over well-managed uniform agricultural fields. The NEECO2 was measured using the eddy covariance method (ECM), and remote sensing signatures were obtained using optical and thermal sensors. A soil-vegetation-atmosphere transfer (SVAT) model was linked with remotely sensed signatures for the simulation of CO2 and water fluxes, as well as biomass, photosynthesis, surface temperatures, and other ecosystem variables. The model was calibrated and validated using an 8-year data set, and the performance of the model was excellent when all necessary input data and parameters were available. However, simulations using the model alone were subject to great uncertainty when some of the important input/parameters such as soil water content were unavailable. Dynamic optimization of parameter/input for the SVAT model using remotely sensed information allows us to infer the target parameters within the model or unknown inputs for the model through iterative optimization procedures. A robust relationship between the leaf area index (LAI) and the normalized difference vegetation index (NDVI) was derived and used for optimization. Our results showed that simulated biomass and NEECO2 agreed well with those measured using destructive sampling and the ECM, respectively. Remotely sensed information can greatly reduce the uncertainty of simulation models by compensating for insufficient availability of data or parameters. This synergistic approach allows the effective use of infrequent and multisource remote sensing data for estimating important ecosystem variables such as biomass growth and ecosystem CO2 flux.

1. Introduction

[2] Terrestrial vegetation plays an important role in global carbon cycle processes. Net primary production (NPP) is the major component of carbon flux at the interface between the atmosphere and terrestrial ecosystems such as agricultural fields, grasslands, and forests [e.g., Roy and Saugier, 2001; Tian et al., 2003]. However, the quantitative dynamics of plant productivity and CO2 flux between the atmosphere and terrestrial ecosystems remains elusive [Keeling et al., 1996; Schimel et al., 2000]. Continuous measurements of CO2 fluxes using eddy covariance or chamber methods are assumed to be the most direct and accurate approaches for the assessment of net ecosystem productivity (NEP). In fact, an increasing number of observational studies using the eddy covariance method (ECM) have been conducted in various terrestrial ecosystems [e.g., Miyata et al., 2000; Carrara et al., 2003; Desai et al., 2005]. Nevertheless, systematic calibrations and methodological evaluations are still under investigation [e.g., Massman and Lee, 2002; Suyker et al., 2004].

[3] Another important issue is the spatial representativeness of such point measurements. Most ecological investigations and assessments are based on site-specific measurements and the assumption that they are representative of the area of interest, but this assumption often results in serious errors [Kicklighter et al., 1994]. Footprint effects on CO2 flux measurements in particular await further experimental and modeling investigations [Schmid, 2002]. Such site-specific observations of CO2 flux must be interpreted at larger geospatial and temporal scales for most geoscientific or assessment purposes [Van Gardingen et al., 1997; Brunsell and Gillies, 2003]. Hence remote sensing and process-based models play crucial roles for these purposes because remote sensing and process-based models allow upscaling and dynamic assessment.

[4] Numerous studies have been conducted to relate remotely sensed signatures to key variables at the interface between ecosystems and the atmosphere, including biomass, leaf area index (LAI), vegetation coverage, evapotranspiration, and the fraction of absorbed photosynthetically active radiation (fAPAR; e.g., see review by Inoue [2003]). The use of fAPAR in a simple mechanistic model [e.g., Monteith, 1977] is a useful approach for the rough assessment of NPP at various scales using remote sensing [e.g., Potter et al., 1993; Maisongrande et al., 1995; Veroustraete et al., 1996; Ruimy et al., 1996; Inoue et al., 1998; Goetz et al., 1999]. Recent studies have shown that microwave backscattering and hyperspectral reflectance signatures may be promising for the estimation of plant productivity and related variables [e.g., Inoue et al., 2002; Verhoef and Bach, 2003; Picard et al., 2003; Haboudane et al., 2004]. One of the most recent approaches uses a hyperspectral reflectance index that is based on photochemical processes in chloroplasts. This may allow the direct assessment of the photosynthetic capacity of vegetation [Peñuelas and Inoue, 2000]. Another important finding is that CO2 flux from bare soil is closely related to remotely sensed surface temperatures [Inoue et al., 2004]. However, few reports have directly related remotely sensed signatures with ecosystem CO2 flux components. Hence it is necessary to extend such findings to a more accurate, consistent, dynamic, and comprehensive assessment of ecosystem variables. On the other hand, process modeling is a powerful approach to incorporate a number of variables and mechanisms into the model structures and to enable continuous simulations [e.g., Stein and Penning de Vries, 1999; Brisson et al., 2003].

[5] Nevertheless, these powerful methods have several inherent limitations in ecological and geophysical applications. Remote sensing observations are basically instantaneous and require biophysical interpretations. Many process models use a number of input variables for which reliable data are often unavailable, especially over large geospatial scales. Parameterization of such models often limits their use [e.g., Sirotenko, 2001; Brisson et al., 2003]. These limitations hamper operational and reliable assessments. Therefore the linkage of these two methods may be promising by facilitating each other and compensating for drawbacks inherent to each method (see a review by Inoue [2003]).

[6] Remotely sensed signatures or biophysical parameters such as LAI and fAPAR estimated from remote sensing data may be used to reparameterize or reinitialize process-based models. Such “assimilation” approach would reduce the model complexity, simplify input requirements, and make the model more operational in the estimation of crop yield [Maas, 1988; Bouman, 1992; Moulin et al., 1998], grassland production [Nouvellon et al., 2001], or evapotranspiration [Ottlé and Vidal-Madjar, 1994; Olioso et al., 1999]. Coupling of radiative transfer model with crop growth model may be useful in this approach [Olioso et al., 1999; Weiss et al., 2001]. Olioso et al. [2005] proposed the linkage of remote sensing signatures in all spectral domains (reflectance, brightness temperature, microwave backscattering) with their radiative transfer models and canopy functioning models. These works suggest that the linkage between remote sensing data with soil-vegetation-atmosphere transfer models (SVAT models) would allow dynamic and accurate estimation of ecosystem CO2 flux, biomass growth and other plant and environmental variables such as soil water content, at the same time.

[7] Thus, in this study, we acquired season long measurements of CO2 flux at fine time resolution over years and investigated the potential of synergy between biophysical/ecophysiological process models and remote sensing for the dynamic assessment of key ecosystem variables such as biomass and CO2 flux.

2. Data Set

2.1. Experimental Site

[8] An experimental study was conducted in agricultural fields at the National Agricultural Research Center in Tsukuba, Japan (36°01′N, 140°07′E, 25 m above sea level), from 1995 to 2002. We used two flat and uniform 4 ha fields that were surrounded by a similar type of cropped field under different management conditions. The soil was a humic volcanic ash Andisol, which belongs to the Hydric Hapludands, and is the major cultivated soil for upland crops in Japan [Hasegawa et al., 1994]. The field capacity and wilting point were estimated to be 0.60 m3 m−3 (−6 kPa) and 0.32 m3 m−3 (−1.5 MPa) for the top 1 m layer, respectively. The mean values of total carbon and nitrogen contents in the topsoil (0–5 cm) were estimated to be 3.7 and 0.31%, respectively for the fields. Each field was subject to the cropping pattern of soybean (Glycine max L. Merr.)–rapeseed (Brassica napus L.)–forage corn (Zea mays L.)–wheat (Triticum spp.) for every 2 years. The average (± standard deviation) of the annual mean air temperature and annual total rainfall were 14.2 ± 0.5°C and 1189 ± 184 mm, respectively, for the 1995–2002 period.

[9] A large data set was obtained through comprehensive long-term experiments in these fields. The data used here are a subset of this larger data set, collected while the experimental fields were cropped in soybean. The average (± standard deviation) air temperature and rainfall for the cropping periods of soybean (June-September) were 23.4 ± 0.7°C and 553 ± 129 mm, respectively. No irrigation water was applied to the fields.

2.2. CO2 Flux Measurements

[10] The CO2 flux from vegetated land surfaces results from photosynthesis by plants and respiration by plants and soil microorganisms. We used the ECM to measure the CO2 flux over the experimental fields (NEECO2). This method has been widely used for the measurement of CO2, water vapor, and sensible heat fluxes over plant canopies [Leuning and Judd, 1996]. We used an open-path eddy covariance system, the main components of which were a three-dimensional sonic anemometer (model DA600, Kaijo Co., Japan), a fast-response infrared gas analyzer (model E009A, Advanet Inc., Japan), and a data logger. The sonic anemometer was used for measuring fluctuations in three components of wind speed and virtual temperature, while fluctuations in CO2 and water vapor concentrations were measured using the gas analyzer. Both sensors had a 10-Hz time resolution and with a path length of 20 cm. The horizontal distance between the open paths of the sonic anemometer and the gas analyzer was 15 cm. The height of the open path was at 1.5 m above the ground. These sensors were installed in the center of the field. It was assumed that measurements at the position was representative of the field since the field size was more than 100 times larger than the height of the sensor and the major part of the footprint was estimated to be within the field [Schmid, 1997].

[11] Data were recorded to a data logger (CR23X, Campbell Scientific Inc., USA), and CO2, H2O, and sensible heat fluxes were calculated from the covariance between vertical wind speed and each component every 20 min. The CO2 flux derived from the covariance was corrected on the basis of the method described by Webb et al. [1980]. Because of inherent limitations of the ECM, data obtained under rainy or calm conditions (average wind speed on a 20-min basis < 0.3 m s−1) were eliminated. Since eddy covariance measurements are subject to unusual fluctuations and/or instrument troubles, we lacked reliable data for some periods. Quality control of flux data is an important issue in the ECM [Miyata et al., 2000], and sometimes a large percentage of flux data has to be eliminated through screening of raw data while interpolation methods are not established. In this case study, we used CO2 flux measurements in 1997 and 1999 to compare simulated and measured flux values. In these 2 years, we had less interruption due to sensor and logger disorders. All CO2 flux data were defined as positive when the direction of flux was from the surface to the atmosphere.

2.3. Remote Sensing Measurements

[12] Spectral reflectance measurements were periodically obtained over the ecosystem using a hyperspectral radiometer and a handheld radiometer. The hyperspectral radiometer (FSFR1000, ASD, USA) covered the wavelength range from 350 to 2500 nm at a 1-nm resolution. The handheld radiometer (Opto-Research, Co. Ltd., Japan) was equipped with seven bands from the visible to short-wave infrared regions (560, 660, 830, 1100, 1200, 1650, and 2200 nm). Field of view of the sensors was 20° and spectral observations were made from a height of 2 m above the canopy at the nadir-looking angle. At least 100 measurements were made each time over different parts of the canopy to obtain representative reflectance spectra for the field. The reflectance factor was calculated relative to the reflectance of a BaSO4 standard panel coated with Kodak Analytical Standard White Reflectance Coating (#6080) to about a 1.0-mm thickness [Inoue et al., 1998]. The white panel was calibrated using a commercial reference board (Spectralon, Labsphere, USA). Reflectance measurements were made for bare soil and soybean canopies during 1995, 1996, 1997, 1998, 1999, and 2001. Similar to the other data sets, reflectance data from 1997 and 1999 were used for validation, while data from the other years were used for deriving relationships between vegetation indices and LAI.

[13] Canopy surface temperatures were measured remotely using three infrared thermometers (Model 4000, Everest Inc., USA) attached to a 3-m pole near the eddy covariance system. The field of view of the sensors was 15°, and the viewing angle of all three sensors was set to be 90° to the ground surface. One of the thermometers was positioned at a height of 3 m and used to measure the average surface temperatures and the other two were attached to booms with adjustable heights to measure the temperature of vegetation only. The emissivity was assumed to be 0.98, since the emissivity of the plants and soil was in the range of 0.95–0.99 [Olioso, 1995]. Data from the infrared thermometers were acquired by the data logger at 1-s intervals and averaged for every 20 min.

2.4. Micrometeorological and Plant Measurements

[14] Air temperature and relative humidity were measured using a probe (HMP45C, Campbell Scientific Inc.) that contained a platinum resistance temperature detector and a capacitive relative humidity sensor. Wind speed was measured using a three-cup anemometer (Model 03001, Campbell Scientific Inc.). Incident and reflected solar radiation were measured with a pyranometer (CM3, Kipp & Zonen, Netherlands). Net radiation was measured using three sensors (Q6, REBS Inc., USA; CG3 and NR-Lite, Kipp & Zonen). Photosynthetically active photon flux density was measured using a quantum sensor (LI-190SB, Li-Cor Inc., USA). Soil heat flux was measured using two heat flux plates (PHF-01, REBS Inc.), and soil temperature was measured using averaging thermocouple probes (TCAV, Campbell Scientific Inc.) that were installed at depths of 5 and 10 cm below the soil surface. Volumetric soil water content was measured using TDR probes (CS615, Campbell Scientific Inc.) for the soil layers at depths of 10, 20, and 30 cm at 20-min intervals during the season. Once during the early growth stage, volumetric soil water content was measured for the depth of 1.0 m using a TDR sensor (TRACE, Soil Moisture Equipment Corp.). These TDR measurements were calibrated on the basis of gravimetric measurements. Micrometeorological variables were measured every 10 s and averaged for every 20 min.

[15] Plant height, wet and dry biomass, and LAI were determined approximately every 10 days by destructive sampling. Ten plant samples were taken each time from the areas more than 50 m away from the eddy covariance sensor so that the effect of plant sampling would be negligible. Sampled plants were divided into stems, green leaves, roots, and dead tissue. The green leaf area was measured using an area meter (AAM8, Hayashi-denkoh Co. Ltd., Japan). The water content of each plant part was determined after desiccation in an oven at 80°C for 48 hours, and was used to calculate dry weights. These plant data were collected for all crops during the experimental period of 1995–2002, and those for soybean crops were used here for calibration and validation of the SVAT model.

3. Process Model and Calibration With Remote Sensing Data

3.1. SVAT Model

[16] We used the soil-vegetation-atmosphere transfer (SVAT) model ISBA-A-gs, described by Calvet et al. [1998]. These authors showed that upon calibration, this model is able to correctly simulate water balance and plant growth. The original version of the model, “Interactions between Soil, Biosphere, and Atmosphere” (ISBA), was developed at Météo-France, to be implemented as a land surface scheme in atmospheric weather forecast models and GCM (general circulation model [Noilhan and Mahfouf, 1996]). The basic parts of the model are described in Appendix A. Details of the model were provided by Calvet et al. [1998], Noilhan and Mahfouf [1996], and Jacobs et al. [1996]. This model solves the surface energy balance and the soil water balance at a 5-min time step. The soil is described by one bulk reservoir corresponding to the maximum root zone (including a thin surface layer and regardless of actual root development). The main surface variables simulated by the model are surface temperature, soil moisture in the root zone, surface soil moisture, and energy fluxes. The model requires meteorological input variables (incoming solar radiation, precipitation, atmospheric pressure, air temperature and humidity, and wind speed), surface albedo, minimum stomatal resistance, LAI, vegetation height, wilting point, and field capacity. The wilting point and field capacity could be estimated from soil texture when they are not available. The ISBA-A-gs has incorporated a physiological submodel to describe photosynthesis, and its coupling with stomatal resistance at the leaf level, by adding three parameters: leaf life expectancy (τM), effective biomass per unit leaf area (α), and mesophyll conductance (gm). The stomatal conductance is expressed as a function of radiation, temperature, vapor pressure deficit, and soil moisture. Introduction of these parameters allows for the simplification of physiological processes and estimation of LAI from biomass production. The computed net vegetation assimilation is used to feed simple submodels for calculation of biomass growth, vegetation cover, and so forth. Thus the model is able to simulate the water budget, energy and mass fluxes (e.g., CO2, sensible and latent heat fluxes), and LAI in response to changes in environmental conditions (e.g., precipitation, irrigation, water storage in the root zone, atmospheric CO2 concentration).

[17] Because the ISBA-A-gs model does not include the soil microbial CO2 flux, we incorporated a simple semiempirical equation between canopy surface temperature and soil microbial CO2 flux into the model based on recent findings by Inoue et al. [2004], who found a close relationship between remotely sensed surface temperature (TIR) and soil surface CO2 flux under bare soil conditions which was assumed to be the soil microbial respiration (SMRCO2). Nevertheless, soil surface CO2 flux is strongly accelerated by turbulent diffusion near the soil surface and it is greatly reduced even by short plants. In fact, a large part of the variability in the SMRCO2 was attributed to wind speed (r2 = 0.6). Therefore, in order to estimate SMRCO2 under vegetated conditions, we derived a semiempirical equation for the lower limit of those scattering points in the TIR – SMRCO2 diagram obtained under bare soil conditions [Inoue et al., 2004].

equation image

This may be used to estimate SMRCO2 under some vegetation cover from the composite surface temperature. We also assumed that SMRCO2 under full plant cover could be estimated from the equation (1) using the canopy surface temperature since the soil surface temperature is close to the composite canopy temperature under vegetated conditions [Moran et al., 1994]. Nevertheless, the SMRCO2-soil surface temperature relationship with and without vegetation is an interesting and important issue to be investigated in detail on the basis of separate measurements of root and microbial CO2 flux components with concurrent measurements of surface temperatures of soil and vegetation as well as wind turbulence over and within a canopy.

[18] The ISBA-A-gs model has been carefully parameterized using various data sets of comprehensive soil, plant, and meteorological measurements, and can be applied to a wide range of ecosystems [Calvet et al., 1998; Olioso et al., 2005]. The majority of parameters may be assumed to be common for different plant types (C3 and C4) and soils [Calvet et al., 1998] or from measurements if available; nevertheless, it is recommended that three parameters, i.e., τM, α, and gm, be refined for each plant species through calibration [Calvet et al., 1998]. Therefore we conducted a calibration for the soybean crops using the data sets obtained in this experiment as described in section 4.

3.2. Synergy Between Remote Sensing Data and the SVAT Model

[19] We attempted to use remotely sensed signatures for tuning of the SVAT model. The SVAT model was further coupled with a semiempirical model to simulate remote sensing signatures or indices using output from the SVAT model. Figure 1 depicts the outline of the structure of the SVAT model and tuning (reparameterization or reinitialization) processes. Thus the coupled model simulates remote sensing indices, as well as ecophysiological variables at each cycle of simulation. The simulated signatures/indices are compared to corresponding measurements by remote sensing; that is, target parameters/inputs in the model are optimally tuned so that the sum of the difference between simulated and measured values is minimized. In principle, the targets for tuning can be chosen among various input/state variables or model parameters, but it is appropriate to choose a few unknown or uncertain parameters or variables. From a practical point of view, the use of remote sensing in this approach may be most effective when a few parameters or inputs with high sensitivity and/or low availability, such as the initial soil water content, are chosen. The optimization procedure was executed using an iterative method based on a simplex algorithm (MATLAB, MathWorks, Inc.).

Figure 1.

Structure of the SVAT model and the procedure for parameterization of the model using remotely sensed signature. The final output results (bottom) can be derived through iterative optimization of parameters or state variables. LAI, leaf area index; gs, stomatal conductance; An, leaf net photosynthesis; Anc, canopy net photosynthesis; Tr, transpiration; ET, evapotranspiration; RS, remote sensing.

[20] In this case study, we attempted to estimate the initial soil water content (θi) and the physiological parameters (τM) through the iterative optimization processes because each θi and τM was one of most important model inputs and parameters, respectively. We used remotely sensed spectral indices (e.g., NDVI) to drive the iterative optimization processes. Details are described in the following section.

4. Results and Discussion

4.1. Calibration of the SVAT Model

[21] Most model parameters can be given from previous studies [Noilhan and Mahfouf, 1996; Jacobs et al., 1996; Calvet et al., 1998] or from measurements if available. Basic input values were latitude and longitude for the modeling area (36°01′N, 140°07′E, respectively), field capacity and wilting point (0.6 m3 m−3 and 0.32 m3 m−3, respectively), plant type (C3), minimum stomatal resistance (50 sm−1), surface albedo (0.18) and surface emissivity (0.98). Several parameters specific to plant conditions such as τM (leaf life expectancy), α (effective biomass per unit leaf area), and gm (mesophyll conductance) must be calibrated to reduce the uncertainty of the simulation.

[22] Therefore we first calibrated gm for the soybean crops, which may be most genetically consistent with each plant species, using measured values of LAI and biomass (dry matter). The calibration of gm was conducted by comparing the time series of simulated and measured biomass values while LAI from measurements were used as a forced variable in this process. This was done in order to break the interactions between LAI production that depends on α and τM and the canopy photosynthesis that depends on mesophyll conductance and LAI.

[23] Figure 2a shows the calibration results for gm using plant data for 6 years (1995–2002, except 1997 and 1999). Data from 1997 and 1999 were used for validation. The value of gm determined for the 6 years was 2.71. The good agreement of measured and simulated biomass suggests that the value of gm was fairly consistent over the years. The value was comparable to gm obtained for soybean crops in Avignon, France (2.60 in the work by Calvet et al. [1998] and 2.87 in the work by Olioso et al. [2005]).

Figure 2.

(a) Calibration results of SVAT model: Comparison between biomass values measured and simulated by the model with the calibrated value of gm. The retrieved value of gm was 2.71. RMSE and r2* indicate the root mean square error, the coefficient of determination adjusted for degree of freedom, respectively. (b) Calibration results of SVAT model: Comparison between LAI values measured and simulated by the SVAT model after calibration of α and τM. The calibrated values of α and τM are shown in Table 1.

[24] In the second step, the other important plant parameters which directly affect LAI values (α and τM) were calibrated using observed and predicted time series of LAI as the value of gm was already determined in the previous step (Figure 2b). Using the common value of gm = 2.71, the values of α and τM were calibrated for all 6 years so that the error between the simulated and measured LAI was minimized. For the 6 calibration years, α = 0.033 and τM = 30.36. We attempted similar parameterization procedures for each of the 6 years separately, and obtained 6-year average values of these parameters (Table 1). Average values of α and τM were 0.0346 (coefficient of variance, CV = 7.8%) and 41.9 (CV = 32.0%), respectively. These results suggest that these parameters may be affected to some extent by year-to-year variability in environmental conditions that are not included explicitly in the model. The variability of τM was much higher than that of α, presumably because it may integrate various physiological effects of environmental conditions. Therefore we parameterized τM for each year, giving a constant value for α (0.033). Results indicated that the ISBA-A-gs model can yield reasonable simulation of LAI and biomass, provided the model is appropriately parameterized.

Table 1. Calibrated Values of Parameters α and τM for Each Yeara
Yearα, kg m−2τM, day
  • a

    gm was 2.71. cv (%) indicates the coefficient of variance. Data were derived through the step shown in Figure 2b.

cv, %7.832.0

4.2. Validation of the Model: Capabilities and Limitations

[25] We validated the model using the independent data sets from 1997 and 1999, which were not used for the calibration. Validation was conducted using parameter values for gm = 2.71 and α = 0.033 obtained through the calibration process described in the previous section, and measured values of initial soil moisture content. Values of τM were determined to be 26.0 for 1997 and 28.1 for 1999, respectively. Figure 3 shows the seasonal comparison of measured and simulated LAI and biomass in the 2 years. Both LAI and biomass in the 2 years were estimated well by the simulation using the SVAT model. Thus we concluded that the performance of the model is excellent, provided that all necessary input data and parameters are available. We also presumed that the ecophysiological processes incorporated in the model work reasonably well because the model has withstood validations from various aspects using a range of data sets including ours [Noilhan and Mahfouf, 1996; Jacobs et al., 1996; Calvet et al., 1998; Calvet and Soussana, 2001].

Figure 3.

Validation results of the calibrated SVAT model by using the data sets from 1997 and 1999. (a) Seasonal comparison between measured and simulated values of LAI and biomass in 1997, respectively. (b) Seasonal comparison between measured and simulated values of LAI and biomass in 1999, respectively.

[26] However, the strong sensitivity of the SVAT model to the initial soil water content has been a common issue for various ecosystems [Calvet et al., 1998, 2004; Calvet and Soussana, 2001]. Figure 4 depicts the seasonal evolution of LAI and volumetric soil water content simulated by the model under different initial soil water conditions. Measured values of soil water content and precipitation are also indicated. The soil water content was for the top 0.3 m soil layer since continuous measurement was not made for the depth of the root zone reservoir (1.0 m). However, the range and seasonal pattern of the measurements were reasonably comparable with those simulation lines. It showed rather sharper response to precipitations than simulation lines since the measured layer was shallower. It is obvious that the simulation results in soybean crops were strongly affected by the initial soil water content (θi). This sensitivity to the initial soil moisture is related to the simple formulations used in ISBA-A-gs for simulating biomass growth and LAI [Calvet et al., 1998]. The decrease of “active” biomass (called “mortality” by Calvet et al. [1998]) depends on the maximum photosynthesis during the previous day. This is directly translated to LAI since LAI is proportional to the simulated “active” biomass. Consequently, if the initial soil moisture is decreased, photosynthesis decreases as a response to soil moisture availability decrease (equation (A10) in Appendix A), mortality increases, and the daily increment of LAI decreases. Along the crop season, photosynthesis, then biomass and LAI, are depressed because of the direct effect of water stress due to the low soil water content and because of the lower LAI. Since the LAI “deficit” is accumulated day after day, this leads to low values of LAI when initial soil moisture is low. It may be possible to modify this behavior by changing the dependence of mortality to photosynthesis in the previous days [e.g., Mougin et al., 1995]. However, we have not yet explored any of these possibilities because good results were usually obtained when we compared ISBA-A-gs model results to experimental data for wheat [Wigneron et al., 1999], corn [Demarty et al., 2004] and soybean [Olioso et al., 2005], and because the implementation of synergy with remote sensing data would help performance of the present model. The major issue in this study was to investigate the effectiveness of linkage between remote sensing and process models with such limitations.

Figure 4.

Simulation results of (a) leaf area index LAI and (b) soil water content at 1 m depth by the SVAT model under different conditions of initial soil water content (θi). Measured value of θi was 0.43 at DOY199. (c) Seasonal change of soil water content at 0.3 m depth and precipitation.

[27] Results in Figure 4 also show that the model can accurately simulate canopy LAI when a realistic initial soil moisture content (measured value = 0.43) was available. Only a slight bias from the real value could results in a large difference in simulation of LAI. Hence it is crucial to obtain accurate estimates of soil water conditions for this SVAT model. In general, soil water content is often required for various models because soil water content near the soil surface is a key variable in the water budget, energy exchange, and plant growth processes. However, soil water content is tedious or difficult to obtain, especially at a necessary geospatial resolution. Because of the limited availability of data on soil water conditions, simulation results from the model alone are subject to great uncertainty. For instance, it is difficult to choose any specific line from those in Figure 4 without inferring the soil water conditions through some relevant method. Therefore the use of remote sensing in conjunction with the SVAT model would be useful to reduce the uncertainty. Although a number of studies have attempted to estimate soil water content using remote sensing of spectral reflectance [e.g., Whiting et al., 2004], microwave scattering [e.g., Wang et al., 2004], or passive microwave signatures [Wigneron et al., 2003], these techniques are basically useful mostly for bare soil conditions, and vegetation cover is always the major source of error for all spectral domains especially when it is dense. Microwave measurements with backscattering models may have potential in the estimation of the soil water content beneath the vegetation cover [e.g., Prevot et al., 1993; Moran et al., 1998]. Nevertheless, remote sensing of soil moisture awaits further investigation, especially for better accuracy and applicability [Wigneron et al., 2003].

4.3. Dynamic Tuning of the SVAT Model Using Remote Sensing Data

[28] It is clear that initial soil water content, θi, is critical for the SVAT model, although it is seldom readily available, especially on a dynamic and wide-area basis. Therefore, in this case study, we attempted to use remotely sensed optical information for optimization of the two important variables, i.e., one environmental input (θi) and one plant parameter (τM), in the SVAT model.

[29] We focused on the relationship between LAI and vegetation indices derived from red and near-infrared wavelength regions because this information is widely available from various remote sensors. We compared the relationship of LAI with typical vegetation indices such as NDVI (normalized difference vegetation index) based on our experimental data (Figure 5).

equation image

where ρ830 and ρ650 are the reflectance at 830 ± 10 nm and 660 ± 5 nm, respectively. Data sets from 1997 and 1999 were not used in deriving the relationship, but were used only for validation, similar to the validation of the SVAT model. The results showed a close and consistent relationship between LAI and NDVI (RMSE = 0.0436, r2* = 0.982) that was consistent and robust over time. Other spectral indices such as SAVI (soil adjusted vegetation index [Huete, 1988]) and WDVI (weighted difference vegetation index [Clevers and Verhoef, 1993]) also showed good relationships (RMSE = 0.0434 and r2* = 0.975 for SAVI; RMSE = 0.0450 and r2* = 0.960 for WDVI).

equation image
equation image

where the coefficients L and c were determined from reflectance of bare soil to be 0.5 and 1.36, respectively. Both SAVI and WDVI were better in linearity but fitted less well than NDVI although SAVI may be better for wider range of soil backgrounds [Huete, 1988]. Thus we used NDVI as a most typical index in this case study. These indices can be replaced with a better index when available. These spectral indices saturate at moderate levels of LAI (∼3.0), but in this approach it is assumed to use a set of a few remote sensing data including some during early growth stage. The tuning process is run using increasing number of remote sensing data along with development of plant growth. Consequently, a set of measured value of NDVI is used to narrow the range of possible time course change of plant growth through the comparison with simulated NDVI values. Figure 6 shows the results of the validation using the independent data sets from 1997 and 1999. Predicted and measured values of NDVI agreed well, and the relationship was assumed to be useful for soybean canopies. Thus we combined this equation with the SVAT model so that measured and simulated NDVI values could be compared in the optimization processes (Figure 1). Consequently, the coupled model was able to simulate NDVI as well as LAI, biomass, and NEECO2.

Figure 5.

Relationship between leaf area index LAI and spectral vegetation index NDVI in soybean canopies derived from measurements in the 4 years: 1995, 1996, 1998 and 2001.

Figure 6.

Validation of the LAI-NDVI relationship shown in Figure 5 using data sets from 1997 and 1999.

[30] Figures 7 and 8show the seasonal comparison between measured and simulated values of biomass and NEECO2 flux after the optimization procedures, in 1997 and 1999, respectively. Data in Figures 7 and 8 were shown to present the typical diurnal changes for early, middle and late growth stages with a range of LAI values. In these trials, the initial soil water content θi and the plant physiological parameter τM were assumed to be unknown because they were not routinely available, but were critical. The squared sum of differences between measured and simulated NDVI was minimized through iterative optimization processes. After minimization of error in the simulated NDVI, θi and τM were determined to be 0.428 and 26.0 for 1997, and 0.450 and 28.1 for 1999, respectively. θi was nearly equal to an independent measurement of 0.44 in 1997 and 0.46 in 1999, respectively. Values of τM were also comparable to those values 26.0 for 1997 and 28.1 for 1999 obtained in the independent validation procedures. These values were automatically retrieved through iteration processes starting with some arbitrary values within the normal range.

Figure 7.

Comparison of seasonal changes of ecosystem CO2 flux, biomass, and canopy surface temperature measured and estimated by the synergy of remote sensing and the process model results in 1997. NEECO2, net ecosystem exchange of CO2; IRT, infrared thermometry; NDVI, normalized difference vegetation index; θi, initial soil water content.

Figure 8.

Comparison of seasonal changes of ecosystem CO2 flux, biomass, and canopy surface temperature measured and estimated by the synergy of remote sensing and the process model results in 1999. Abbreviations are the same as in Figure 7.

[31] Simulated seasonal changes in dry matter production agreed well with those obtained by destructive measurements. Simulated values of NEECO2 were in good agreement with independent measurements of NEECO2 by ECM. Values of the coefficient of determination (r2*) for NEECO2 ranged from 0.7 (RMSE = 0.042) to 0.94 (RMSE = 0.080). Those for surface temperature ranged from 0.85 (RMSE = 2.17) to 0.98 (RMSE = 0.77). Those values for the pooled data from the two seasons were 0.88 (RMSE = 0.117, n = 6668) for NEECO2, and 0.93 (RMSE = 1.49; n = 7719) for surface temperature, respectively. The overall trends and statistical indicators suggest the soundness of the simulation results, i.e., this approach. However, there were some overestimates or underestimates during the periods, which may be explained by the simple structure of the model and the effects of unconsidered microclimate conditions. The simulated canopy surface temperatures were also comparable to those measured using the infrared thermometer. Simulated canopy temperature, especially during periods of greater vegetation cover, fit well with the measurements. This suggests that remotely sensed surface temperature could be used for driving the recalibration/reinitialization method in place of, or together with the optical remote sensing measurements. Nevertheless, some temporal shift was found between measured and simulated surface temperatures even though diurnal pattern was similar. It may be due to the difference in time constants of thermal response to the changing environment, which is not incorporated in the SVAT model. Further investigation on the behavior of the composite surface temperature is needed.

[32] It was obvious that the uncertainty using only the simulation was too large to assess seasonal changes without information on the soil water condition (θi) and/or plant physiological status (τM). Remotely sensed information can greatly reduce this uncertainty without any direct information on soil water and physiological conditions. In conclusion, remote sensing will play a significant role in reducing uncertainty of the dynamic simulation of ecosystems by compensating for critical data. In this case study, we used some intermittent measurements of remote sensing data (NDVI) during the growing season, which was enough for this iterative optimization. An effective use of infrequent measurements is one of the important advantages of this approach. Nevertheless, timing of the remote sensing observation is also important since the sensitivity of spectral indices such as NDVI change with growth. It is required to utilize at least a few measurements before the saturation of NDVI for better accuracy.

5. Concluding Remarks

[33] The calibration and validation results using an 8-year data set showed that the performance of the SVAT model was excellent when all necessary input data and parameters were available. However, simulations using the model alone were subject to great uncertainty when some of the important input/parameters such as soil water content were unavailable. Dynamic optimization of parameter/input for the SVAT model using remotely sensed information allowed us to infer the target parameters within the model or unknown inputs for the model through iterative optimization procedures. A robust relationship between the leaf area index (LAI) and the normalized difference vegetation index (NDVI) could be used for optimization. Simulated biomass and NEECO2 agreed well with those measured using destructive sampling and the ECM, respectively. This study clearly demonstrated that the combined use of remote sensing and SVAT modeling was effective for estimating important ecosystem variables such as plant biomass and ecosystem CO2 flux. The linkage, i.e., the iterative parameterization/initialization of process-based models using remotely sensed signatures, may be one of useful approaches for dynamic assessment of ecosystem variables, since it allows for the use of infrequent and multisource remote sensing data.

[34] Radiative transfer models in the microwave domain, as well as optical and thermal domains, may also be linked with the SVAT model to extend the applicability [Olioso et al., 2005]. One of the interesting issues to be investigated is the relative usefulness of optical, thermal, and microwave signatures in such linkage. This may be critical, especially in using multisource signatures together in the approach. Another issue is the applicability of this approach to upscaling of those ecosystem variables in various types of terrestrial ecosystems. In general, applicability and accuracy of such approaches largely depend on the availability of spatial data at necessary resolutions and on the accuracy of calibration of the ecosystem models. Therefore, at present, the use of this approach may be rather limited to a uniform ecosystem such as an agricultural field, but recent advances in meteorological information network and high temporal/spatial resolution remote sensing data will allow operational applications of combined use of process model and remote sensing data [e.g., Doraiswamy et al., 2004]. It is strongly recommended to include some ground-based, airborne, and/or spaceborne remote sensing measurements in ecological and environmental field experiments to extend the point or instantaneous measurements to geospatial and dynamic evaluations.

Appendix A:: Basic Equations for Energy Balance and Photosynthesis in the SVAT Model

A1. Energy and Water Balance

[35] Energy balance for a canopy is expressed by the following equation,

equation image

where Rn is the net radiation; RG the incoming solar radiation; α the surface albedo; ɛt the emissivity; RA the atmospheric infrared radiation; σ the Stefan-Boltzmann constant; Ts the surface temperature; H the sensible heat flux; LE the latent heat flux; and G the ground heat flux, respectively.

[36] The turbulent fluxes are calculated by the general aerodynamic formulae;

equation image
equation image
equation image
equation image
equation image

where Cp is the specific heat of air; ρa, Va, and Ta are respectively the air density, the wind speed, and the temperature at the lowest atmospheric level; L is the latent heat of vaporization; Eg the evaporation from soil surface; Ev the transpiration; vc the vegetation cover; hu the relative humidity at the ground surface; qsat the saturated specific humidity, qa the atmospheric specific humidity; hv, the Halstead coefficient; wr mean volumetric water content; the Pg precipitation; ρw density of liquid water; and d the depth of soil layer, respectively.

A2. Photosynthesis

[37] The leaf net photosynthesis under light-saturating condition is expressed as a function of maximum net photosynthesis Am,max, mesophyll conductance gm, leaf internal CO2 concentration Ci, and the compensation point Γ.

equation image

The leaf net photosynthesis within a canopy is affected by the light intensity Ia as,

equation image

where ɛ is the initial quantum use efficiency, and Rd the leaf respiration, respectively. The canopy net photosynthesis is expressed by the following equation as integration of (A8) to the depth of the canopy.

equation image

where LAI is the leaf area index, z the height from the soil surface, and h the canopy height.

equation image

where g is the mesophyll conductance under the normalized soil moisture θn; wwilt and wfc are soil water content at wilting point and field capacity, respectively.


[38] The authors wish to thank C. J. Calvet for his kind help in use of the SVAT model. This work was supported partly by the Global Environment Research Fund, Ministry of Environment, and partly by the Ministry of Education, Culture, Sports, Science and Technology in Japan.