Remote Sensing of Environment 82 (2002) 188 – 197 www.elsevier.com/locate/rse
Designing a spectral index to estimate vegetation water content from remote sensing data: Part 1 Theoretical approach Pietro Ceccato a,*, Nadine Gobron a,1, Ste´phane Flasse b,2, Bernard Pinty a,3, Stefano Tarantola c,4 a
Global Vegetation Monitoring Unit, Institute for Environment and Sustainability, Joint Research Centre of the European Commission, TP 440, 21020 Ispra, Varese, Italy b Natural Resources Management Department/Natural Resources Institute, Medway University Campus, Central Avenue, Chatham Maritime, Kent ME4 4TB, UK c Technological and Economic Risk Management Unit, Institute for the Protection and Security of the Citizen, Joint Research Centre of the European Commission, TP 361, 21020 Ispra, Varese, Italy Received 29 May 2001; received in revised form 19 January 2002; accepted 16 March 2002
Abstract This paper describes the methodology used to create a spectral index to retrieve vegetation water content from remotely sensed data in the solar spectrum domain. A global sensitivity analysis (GSA) using radiative transfer models is used to understand and quantify vegetation water content effects on the signal measured at three levels: leaf, canopy, and atmosphere. An index is then created that optimises retrieval of vegetation water content (in terms of water quantity per unit area at canopy level) and minimises perturbing effects of geophysical and atmospheric effects. The new index, optimised for the new SPOT-VEGETATION sensor, is presented as an example. Limitations and robustness of the index are also discussed. D 2002 Elsevier Science Inc. All rights reserved.
1. Introduction Estimation of vegetation water content is central to the understanding of biomass burning processes. Recently, an international fire team, part of the project for Global Observation of Forest Cover, has recognised the necessity to improve satellite technologies and methods to generate accurate, global wildland fire fuel maps (Dull & Singh, 2000). One important factor of the maps is the fuel moisture used to assess the risks of fire occurrence and burning
* Corresponding author. Current address: AGPP Room C798, FAO (UN), Viale della Terme di Caracalla, 00100 Rome, Italy. Fax: +39-0657055271. E-mail addresses:
[email protected] (P. Ceccato),
[email protected] (N. Gobron),
[email protected] (S. Flasse),
[email protected] (B. Pinty),
[email protected] (S. Tarantola). 1 Fax: + 39-0332-789073. 2 Currently with Flasse Consulting, 3, Sycamore Crescent, Maidstone ME16 OAG, UK. Fax: + 1-5308-845-626. 3 Fax: + 39-0332-785733. 4 Fax: + 39-0032-785733.
efficiency. Research on assessing vegetation moisture content has already been carried out using sensors working in the solar spectrum, thermal, and radar domains. However, the methods developed and tested over small regions do not allow the retrieval of vegetation water content, operationally, over different ecosystems, from local to global scale (see Ceccato, Flasse, Tarantola, Jacquemoud, & Gre´goire, 2001, for a more detailed discussion). Various methods can be used to extract specific information via remote sensing. One method is to invert models describing the physical processes that determine the measurements (Verstraete & Pinty, 1996). By constraining the inversion process, the joint use of measured and modelled angular fields of the scattered light reflected by the different components of the Earth yields more accurate and reliable retrievals (Diner et al., 1999). The arrival of the new Multiangle Imaging SpectroRadiometer (MISR) provides measurements from cameras pointing in nine different directions. These angular field measurements will therefore increase the precision of retrievals. However, for the time being, simpler methods using satellites looking straight down are still required to access the desired information. These methods
0034-4257/02/$ - see front matter D 2002 Elsevier Science Inc. All rights reserved. PII: S 0 0 3 4 - 4 2 5 7 ( 0 2 ) 0 0 0 3 7 - 8
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
are based on the development of indices and allow the rapid processing of large amounts of remote sensing data. Spectral indices have been and are still widely used to retrieve information on vegetation biophysical properties. However, the creation and use of spectral indices are often performed with empirical methods. Any significant improvement of the indices, for instance, through corrections using the bidirectional reflectance distribution function (BRDF), enhances their utility and accuracy (Asner, Braswell, Schimel, & Wessman, 1998). In the solar spectrum domain, several attempts to estimate vegetation water content on the basis of satellite remote sensing data have relied on empirical relationships between vegetation moisture content and values of particular spectral indices (mostly the Normalized Difference Vegetation Index, NDVI) computed from spectral reflectance (Ceccato et al., 2001; Tucker, 1980). At leaf level, research using laboratory measurements and model simulations has shown that the use of reflectance measured in the optical domain can provide an estimation of leaf water content in terms of equivalent water thickness (EWT) expressed in quantity of water per unit area (g cm 2) and not in terms of moisture content expressed in quantity of water per quantity of fresh or dry matter (%) (Ceccato et al., 2001). In addition, the authors showed that both the shortwave infrared (SWIR) and the near infrared (NIR) wavelength ranges are necessary for retrieving EWT at leaf level. At canopy level, several authors proposed relationships between vegetation water content and spectral indices (Downing, Carter, Holladay, & Cibula, 1993; Hunt & Rock, 1989). Using the hyperspectral radiometer Airborne Visible InfraRed Imaging Spectrometer (AVIRIS), Roberts, Green, and Adams (1997) and Ustin et al. (1998) studied the possibility to retrieve canopy vegetation water content. In addition, Dawson, Curran, North, and Plummer (1999) showed that the water index (WI) proposed by Pen˜uelas, Filella, Biel, Serrano, and Save´ (1993) and the normalized difference water index (NDWI) proposed by Gao (1996) were related to the quantity of water per unit area in the canopy. The authors determined the canopy water content by scaling the foliar water content (FWC, %) with the specific leaf area (SLA), leaf area index (LAI), and the percent canopy cover for a specific forest canopy. However, their method requires knowledge on SLA, which varies according to species and phenological status. When creating the NDWI, Gao suggested that the research be further developed, in particular using radiative transfer modelling, in order to understand this index well enough to use it with the new generation of satellite instruments such as MODIS and SPOT-VEGETATION. With the arrival of these new sensors carrying channels in the SWIR and NIR wavelength ranges, it becomes interesting, from a theoretical and practical point of view, to determine the limits of feasibility to retrieve reliably and accurately vegetation water content from remotely sensed observations. The purpose of this paper is to expand the understanding of how variations in vegetation water content affect the
189
signal measured at the satellite level and to create an index that is optimised for retrieving vegetation water content. The index aims to provide an operational method for quantitatively retrieving vegetation water content from local to global scale in a rapid and reliable fashion. In Section 2, a study is performed using a global sensitivity analysis (GSA) to identify the requirements prior to the development of the index. The GSA identifies the importance of vegetation water content variation, as well as other factors, on reflectance measured at the satellite level. It also identifies which wavelength regions can provide a solution to retrieve vegetation water content. In Section 3, this information is used to create an index that maximises the accuracy of water content retrieval while minimising the effects of atmospheric scattering and absorption, as well as temporal and spatial variations in the geometry of illumination and observation. The characteristics of the new sensor SPOT-VEGETATION (Mayaux, Gond, & Bartholome´, 2000) are chosen to show how the method can be applied. Other sensors having wavelength bands similar to those described in the requirements could also potentially be used.
2. Identifying the requirements Before developing the index, it is important to understand how vegetation water and other factors affect the signal measured at satellite level and which wavelengths can provide a solution to the problem of retrieving vegetation water content from the satellite measurement. Since reflectance values measured at the satellite level result from the transfer of radiation through a system of three media, leaf, vegetation coupled with soil, and atmosphere, it is thus necessary to study the effects of the factors at each of these three levels. This study was done using a GSA. The GSA was performed using an extension of the Fourier Amplitude Sensitivity Test (EFAST) originally developed by Cukier et al. (Cukier, Fortuin, Schuler, Petschek, & Schaibly, 1973; Chukier, Levine, & Schuler, 1978) and further extended by Saltelli, Tarantola, and Chan (1999). EFAST studies how variation in the output of a model can be apportioned, qualitatively or quantitatively, to different sources of variation (Saltelli, 2000). The basis of the EFAST approach is a parametric transformation that enables reducing multidimensional integrals over the space of the input factors to one-dimensional quadratures, via a search curve that scans the whole input space. The scanning is done so that each axis of the factor space is explored with a different frequency. This approach allows the definition of a structured and restricted set of simulations for which all input factors vary simultaneously. A Fourier decomposition is then used to obtain the fractional contribution of the individual input factors to the variance of the model prediction (Campolongo, Saltelli, Sørensen, & Tarantola, 2000). EFAST provides two sets
190
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
Table 1 Range of variations for PROSPECT model Parameters Meaning EWT N Cm Ca + b
leaf leaf leaf leaf
equivalent water thickness internal structure dry matter content chlorophyll a + b content
Unit
Range of variation 2
g cm – g cm 2 Ag cm 2
[0.0001 – 0.07] [1 – 4] [0.001 – 0.040] [0.01 – 60.0]
of indices: first-order and total indices. The first-order indices give the additive effect of the corresponding factors while the total indices are overall measures of importance and, as such, take into account the effects of the interactions of each factor with the others. To demonstrate the GSA, a three-factor case is taken for illustrative purposes. The variance of the output of interest is decomposed via EFAST as followed: V ¼ V1 þ V2 þ V3 þ V12 þ V13 þ V23 þ V123
ð1Þ
where V is the variance of the output of interest, V1 is the variance of input factor 1, V12 is the variance of interaction between factors 1 and 2, and V123 is the variance of interaction among factors 1, 2, and 3. The first-order partial variances (i.e., V1, V2, and V3) that can be computed via EFAST represent the contribution of one single source (an input factor) to the model output (in terms of variance). The ratio between the partial input variance and the total variance is called first-order index, S1 ¼ V1 =V
ð2Þ
where S1 is the first-order index for input factor 1, V1 is the variance of input factor 1, and V is the variance of the output (see Eq. (1)). The first-order index (see Eq. (2)) for a given input factor represents the percentage of the output variance that is accounted for by that input factor. The second and higher
order terms (i.e., V12 + V13 + V23 + V123) represent the importance of two-way or higher-way interaction between factors. If there is no interaction among the factors in the model, these terms are zero. The total index is a measure of the variance due to the individual factor and all interactions among the input factors. As such, the total index (Eq. (3)) quantifies the degree of additivity of the model. In equation form, the total index is measured as follows: S ¼ ðV1 þ V12 þ V13 þ V123 Þ=V
ð3Þ
where S is the total index, V1 is the variance of input factor 1, and the other terms are the variances of interactions between factor 1 and the other factors 2 and 3. In decomposing variation to the different order indices, the GSA allow us to isolate the importance of any single or any combination of input factors. Ideally, the GSA should be performed on a broad spectrum of data to yield results that are globally applicable. Available laboratory and field measurements are not exhaustive enough to provide reflectance measurements for different types of vegetation around the world. These measurements are also expensive to carry out. However, controlled simulations and experiments can be easily conducted using radiative transfer models at leaf, canopy, and atmospheric levels. Because of this, the GSA is performed using reflectance values simulated with radiative transfer models at each of the three levels in the spectral domain between wavelengths of 400 and 2500 nm. 2.1. Leaf level A preliminary GSA, based on EFAST, was performed by Ceccato et al. (2001) using the new version of the PROSPECT model (Jacquemoud, Bacour, Poilve´, & Frangi, 2000). The GSA was performed on two specific wave-
Fig. 1. Sensitivity analysis of leaf spectral reflectance to leaf characteristics. The y-axis provides the percentage for which each variable accounts for reflectance variation. The filter functions for SPOT-VEGETATION (SPOT-VGT) bands are plotted on the same y-axis.
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197 Table 2 Range of input factors for top-of-canopy level
191
2.2. Canopy level
Factors
Unit
Range of variation
Distribution
Reflectance at 820 nm Reflectance at 1600 nm Transmittance at 820 nm Transmittance at 1600 nm Satellite zenith angle LAI Sun zenith angle Leaf orientation
% % % % j m2 m 2 j –
uniform uniform uniform uniform uniform uniform uniform uniform
Canopy height Leaf radius
m m
[35, 60] [20, 50] [15, 40] [10, 35] [ 80j, 80j] [1, 9] [0j, 60j] planophile, erectophile [0.5, 10.0] [0.01, 0.20]
uniform uniform
lengths, 820 and 1600 nm. In the present study, the GSA has been extended to the whole spectrum between 400 and 2500 nm using a range of values (Table 1) as described in the article by Ceccato et al. (2001). Fig. 1 illustrates the results of the sensitivity analysis performed on the set of reflectance values simulated with PROSPECT. The firstorder indices, expressed in percentage, are plotted on the yaxis for each leaf factor: Ca + b (chlorophyll a + b concentration), EWT, Cm (dry matter content), and N (internal structure parameter). The sum of the first-order indices is close to 100%. This means that uncertainty in the output is not driven by interaction among the factors in the PROSPECT model. The results show that between 400 and 710 nm, reflectance values are primarily influenced by the chlorophyll content and secondly by the N factor. In the NIR region, variations in reflectance values are exclusively influenced by N and Cm, with N having the greatest influence (70 – 80% of reflectance variations are due to N) and Cm accounting for the remaining variations (30 – 20%). In the SWIR region, the EWT factor has an influence on the leaf reflectance that varies according to the wavelength considered. However, the N and Cm factors also significantly affect reflectance values within this range. These GSA results show that the SWIR region alone is not enough to retrieve EWT. Additional information is required on the factors N and Cm. The uncertainty brought by N and Cm in retrieving EWT using reflectance values in the SWIR region can be minimised by using a wavelength in the region affected only by these two factors. The NIR region is therefore required in combination with SWIR to retrieve EWT. The SPOT-VEGETATION sensor with its two wavelength bands in, respectively, the NIR and SWIR (see the filter function of the SPOT-VEGETATION in Fig. 1) is a potential candidate. Other sensors with similar wavelength bands (such as MODIS) could also be used to retrieve vegetation water. More generally, Fig. 1 could also be used to select specific wavelength bands where reflectance values are more sensitive to the EWT rather the N and Cm factors. The use of a hyperspectral radiometer such as the AVIRIS would allow the selection of such specific regions.
Biophysical sources of variability in canopy reflectance and bidirectional reflectance function (BRF) variations due to observing geometry have already been studied by several authors (Asner, 1998; Gastellu-Etchegorry et al., 1999; Jacquemoud, 1993). However, as stated by Bacour, Jacquemoud, Tourbier, Dechambre, and Frangi (in press), the sensitivity analysis methodology used to estimate the importance of each factor faces problems due to the design of simulations that do not explicitly report possible interaction between factors. To complement the study, a GSA, based on EFAST, was also performed to quantify the relative importance of biophysical and observing geometry factors as well as their interactions. The GSA was performed using the output reflectance at the top-of-canopy simulated with the semidiscrete model developed by Gobron, Pinty, Verstraete, and Govaerts (1997). Various ranges of input factors were used to simulate the radiance field (Table 2). Minimum and maximum values selected for leaf reflectance and transmittance were obtained from those simulated with PROSPECT in the previous section. The ranges of values characterising the canopy properties were selected to represent different ecosystems. The sun and satellite zenith angles were made variable in the principal plane, where the maximum variance in reflectance and the phenomenon of hot spots (high reflectances in the backward scattering direction) are observed. The GSA at top-of-canopy was performed only for specific wavelengths of interest, as opposed to the study at leaf level that investigated the whole solar spectrum. This limitation was made due to the significant computation time for performing the simulations and the sensitivity analysis. Respectively, wavelengths at 820 and 1600 nm were selected to represent the NIR and SWIR regions covered by the SPOT-VEGETATION sensor. Soil perturbing effects on reflectance at top-of-canopy were taken into account by using five different soil types ranging from dark to bright. Table 3 shows the soil reflectance values that were chosen from Price (1995). For this study, the effects of soil water content variations on soil reflectance values were not taken into account. For the purposes of this simulation experiment, this is appropriate because the goal is isolation of vegetation water content.
Table 3 Dry soil reflectance values used for each wavelength Soil reflectance (%)
Dark
Medium dark
Medium
Medium bright
Bright
At 820 nm At 1600 nm
3.33 3.60
15.58 20.65
28.70 35.79
35.05 43.66
54.12 54.90
Values were extracted from measurements performed by Price (1995).
192
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
Table 4 Sensitivity analysis results at canopy level: (a) first-order and (b) total indices Factors
Leaf reflectance Satellite zenith angle Leaf transmittance LAI Sun zenith angle Leaf orientation Soil Canopy height Leaf radius
First-order indices (%)
Total indices (%)
At 820 nm
At 1600 nm
At 820 nm
At 1600 nm
28.06 16.94 18.89 3.46 1.07 8.63 0.95 0.70 0.71
41.33 19.44 10.46 0.71 2.05 4.10 0.54 0.75 0.63
40.67 40.14 24.86 17.36 11.14 14.24 11.89 7.35 6.70
55.43 48.15 14.87 13.65 18.23 14.26 9.5 6.04 7.18
From the range of values listed above, an input sample of 945 combinations of canopy and soil factors was provided by EFAST and used to simulate 945 reflectance values for each wavelength using the semidiscrete model. The first-order indices of EFAST (Table 4a) show that, at each wavelength, the sum of all the first-order indices is approximately 80%. This means that about 80% of the uncertainty in the model output is explained by the factors singularly. The remaining 20% is explained by interactions between the factors. Therefore, the EFAST total indices are provided (Table 4b) to account for the additive effect of each factor and their interactions with the others. In Table 4b, the ranking of the factors affecting the topof-canopy reflectance show that reflectance variations measured in SWIR and NIR are primarily influenced by leaf reflectance. A variation in the reflectance measured at topof-canopy is therefore mainly the result of variations in the factors at leaf level (N, Cm, EWT). The other factors significantly influencing the reflectance values are the satellite zenith angle, leaf orientation, LAI, the sun zenith angle, and soil types. Compared to the other factors, the canopy height and leaf radius are less important. These two factors are known to lead to the hot spot phenomenon corresponding to an increase in the reflectance field particularly noticeable for observation directions close to the illumination direction (Gobron et al., 1997). A specific GSA confirmed that their importance increased when those specific conditions are applied. However, those observing conditions are rarely observed with SPOT-VEGETATION. Additionally, simulation results showed that the reflectance values simulated with LAI greater than 5 are invariable at 1.6 Am, confirming the findings of Lillesaeter (1982). 2.3. Atmospheric level Atmospheric gases and aerosols contribute to absorption and scattering of direct sunlight and sunlight reflected from the Earth’s surface (Kaufman, 1989). The main components of atmospheric gases are nitrogen, oxygen, carbon dioxide, ozone, and water vapour. The term ‘‘atmospheric aerosols’’
refers to the liquid and solid matter suspended in air. In the spectral region defined by the NIR and SWIR, only the water vapour and aerosols affect the signal retrieved at satellite level. It is well known that aerosol scattering effects are very strongly dependent on wavelength, especially at shorter wavelengths (Govaerts, Verstraete, Pinty, & Gobron, 1999). Hence, measurements in the blue region of the solar spectrum will provide values much more sensitive to atmospheric scattering than those at longer wavelengths (Gobron, Pinty, Verstraete, & Govaerts, 1999). Similarly, the NIR region is more sensitive to atmospheric scattering than the SWIR region. To minimise aerosol effects, therefore, the blue channel can be used to rectify the other channels as suggested by Kaufman and Tanre´ (1992) for the RED and by Govaerts et al. (1999) for the NIR. The influence of water vapour is less documented for the SPOT-VEGETATION wavelength bands. A GSA was performed to understand the influence of water vapour using the 6S model (Second Simulation of the Satellite Signal in the Solar Spectrum) from Vermote, Tanre´, Deuze´, Herman, and Morcrette (1997). Aerosol optical thickness at 550 nm from 0.00 to 1.00 and water vapour content values from 1.00 to 5.00 g cm 2 were used to simulate different atmospheres characterising an evolution from clear to a thick atmosphere. From an input sample of 194 combinations of aerosols and water content, reflectance values were simulated at the three wavelength regions representing the SPOT-VEGETATION BLUE, NIR, and SWIR channels. The first-order indices, expressed in percentage (Table 5) showed that the effect of water vapour is only significant (responsible for 11.12% of variations) in the NIR spectral band of SPOT-VEGETATION. In addition, the sum of the first-order indices, near to 100% indicated that uncertainty in the output is not driven by interaction among the factors in the 6S model. Water vapour perturbation could be avoided by selecting carefully the spectral location of narrow bands or by using a band sensitive to water vapour to correct the NIR in the same manner used for aerosols. However, these two solutions are not possible with the SPOT-VEGETATION sensor. Reduction of water vapour perturbation would be possible with a hyperspectral radiometer such as AVIRIS or a sensor carrying a spectral band sensitive to the water vapour such as MODIS. Should this method be applied using another sensor that has these capabilities, water vapour perturbations could be minimised further.
Table 5 Sensitivity analysis results at the top-of-atmosphere for the BLUE, NIR, and SWIR SPOT-VEGETATION channels Factors
Aerosols Water vapour
Percent variation (%) BLUE
NIR
SWIR
98.73 0.00
87.61 11.12
97.50 1.25
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
Section 3 of this article describes how, using the NIR, SWIR, and BLUE channels from SPOT-VEGETATION, a method can be designed to retrieve vegetation water content.
3. Designing the index The theory behind the design of optimal spectral indices has been taken from the one described by Gobron et al. (1999), Govaerts et al. (1999), and Verstraete and Pinty (1996). A rational procedure to design the optimal index is to build on radiation transfer models of the associated leaf surface– atmosphere system to simulate sensor-like data sets over various representative land surfaces (Verstraete & Pinty, 1996). This approach defines a large number of simulated radiance fields, which can be sampled by a virtual instrument similar to the actual one in terms of spectral and angular observing sampling scheme (Gobron, Pinty, Verstraete, & Widlowski, 2000). The design of optimal spectral indices is based on a two-step procedure where the spectral radiances simulated in the NIR band are, first, rectified in order to ensure the ‘‘decontamination’’ from atmospheric and angular effects. Second, the rectified NIR simulated values are combined with simulated spectral radiances in the SWIR to generate the optimal index formulae. 3.1. Simulation of reflectance values in the rectified NIR – SWIR spectral space To quantify the range of variations produced by vegetation water content in the NIR and SWIR spectral space, reflectance values were simulated using the asso-
193
ciated three radiation transfer models (PROSPECT –Semidiscrete – 6S). The results of the sensitivity analysis enabled the selection of specific value ranges for the biophysical and observation factors that were judged to be most illustrative (Table 6). In addition, care was taken to select a wide sampling of biophysical and observation values so as to generate a robust index that would be applicable for different ecosystems around the world. At leaf level, four spectral profiles were simulated using four different EWT values extracted from the logarithmic scale used in the sensitivity analysis. The four EWT values were selected to provide data that were equally distributed across the logarithmic range of 0.0001 –0.07. The other three leaf factors were kept constant since either the combination of NIR and SWIR minimise their effects (in the case of dry matter and internal structure) or the effect is null (in the case of chlorophyll). At canopy level, reflectance values were simulated using a LAI ranging from 1 to 5, for the same reasons as in the sensitivity analysis. Values for the leaf angle distribution were taken at two opposite profiles, erectophile and planophile. These values were chosen because they significantly affect reflectance variations as shown in the sensitivity analysis. The height of canopy and leaf radius were kept constant since the sensitivity analysis showed that their variations have little impact on reflectance variations (except for the hot spot phenomenon that is rarely observed with SPOT-VEGETATION). Two soil spectra of contrasting brightness were selected from the Price database to provide an indication of how extreme differences affect the reflectance values. To account for the scattering and gaseous transmittances due to the atmosphere, sky conditions characterised by the
Table 6 Biophysical scenarios, illumination, and observation geometries to simulate the reflectance fields Medium Leaf model (PROSPECT) (Jacquemoud et al., 2000)
Vegetation model (Gobron et al., 1997)
Soil database (Price, 1995)
Atmosphere 6S model (Vermote et al., 1997)
Variable EWT Chlorophyll a + b Dry matter Internal structure LAI Leaf angle distribution Height of canopy Leaf radius Solar zenith angle Solar azimuth angle Sensor zenith angle Sensor azimuth angle Soil reflectance At 460 nm At 820 nm At 1640 nm Aerosol optical thickness at 550 nm Water vapour
Unit
Range of values 2
g cm Ag cm 2 g cm 2
% % % –
[0.0002, 0.0038, 0.0163, 0.07] 33.0 0.01 2 [1,2,3,4, 5] erectophile, planophile 2 0.10 20 30 [0, 10, 20, 30, 40, 50, 60, 70, 80] [0, 90, 180, 270] Dark soil 5.56 8.49 7.67 [0.05, 0.3, 0.8]
g cm 2
[1.00, 4.00]
– – m m j j j j
Bright soil 7.90 20.61 23.87
194
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
US62 standard atmospheric vertical profile and continental aerosol types were assumed. The aerosol optical thickness at 550 nm assumed three values 0.05, 0.3, and 0.8, characterising an evolution from a clear (0.05) to a thick (0.8) atmosphere. In addition to these biophysical and aerosol values, Table 6 also describes the observation scenarios chosen to specify the conditions under which the selected scenes will be illuminated by the sun and observed by the instrument. Rectification of the NIR band for the SPOT-VEGETATION sensor was performed combining the BLUE and NIR channels to generate a ‘‘rectified NIR channel’’ following the method and parameter values used by Gobron et al. (2000). Reflectance values corresponding to the SPOTVEGETATION BLUE channel were simulated using the same scenarios described in the previous section and the resulting values were combined with the NIR-simulated values. The output of the procedure was a rectified NIR channel whose values were decontaminated from atmospheric effects. 3.2. Optimisation of the spectral index The values simulated for the new rectified NIR channel and the SWIR were combined to visualise the range of variations taken by the different scenarios (Fig. 2). Fig. 2 shows clear groupings implying a displacement in the SWIR reflectance values from high to low values when EWT increases. However, within each group represented by an
EWT value, variations due to the LAI provide another level of groupings. This second level of groupings shows that, at canopy level, the use of a combination of EWT and LAI can indicate with further precision a displacement in the rectified NIR – SWIR spectral space. The combination of EWT and LAI therefore can represent a new state variable that corresponds to a quantity of water per unit area in the canopy (EWTcanopy): EWTcanopy ¼ LAI EWT
ð4Þ
where EWTcanopy is expressed in grams per square meter (g m 2), LAI in square meter per square meter (m2 m 2), and EWT in grams per square meter (g m 2). EWT is defined as a hypothetical thickness of a single layer of water averaged over the whole leaf area (Danson, Steven, Mathus, & Clark, 1992) and LAI is the area of leaf surface per unit of soil surface. Therefore, the new variable EWTcanopy (EWT at canopy level, Eq. (4)) is defined as a hypothetical thickness of a single layer of water multiplied by the number of layers determined by the LAI. EWTcanopy thus represents a quantity of water per unit area at canopy level. For a LAI value equal to 1, the results of the simulation show a greater dispersion than the ones observed for a LAI greater than 2. This is due to soil effects. The two reflectance values selected to represent bright and dark soils are affecting the dispersion of the simulated values with an increasing effect from low to high EWTcanopy values.
Fig. 2. Scatter diagram of reflectance values simulated at the rectified NIR channel versus SWIR channel. EWT values are denoted by the symbols and the LAI values by the gray gradient.
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
195
Fig. 3. GVMI isolines in the SWIR and rectified NIR spectral space. Regression lines displayed were computed for each EWTcanopy value.
Based on the rectified NIR and the SWIR, a spectral index is defined to estimate vegetation water content in terms of EWTcanopy. To create the new index, called Global Vegetation Moisture Index (GVMI), regression lines were computed and displayed in the spectral space (Fig. 3) for each EWTcanopy value. Regression line displacement therefore represents EWTcanopy displacement. Next, isolines representing the index were fitted to the regression lines in such a way that the isolines were parallel to the regression lines and therefore orthogonal to the EWTcanopy displacement vector (Fig. 3).
Eq. (5) is the mathematical equation that represents the GVMI isolines: GVMI ¼
ðNIRrect þ 0:1Þ ðSWIR þ 0:02Þ ðNIRrect þ 0:1Þ þ ðSWIR þ 0:02Þ
ð5Þ
with values of GVMI varying between 0.0 and 0.9. Fig. 3 also shows that the regression lines marked LAI 1 are not parallel to the index isolines. These regression lines correspond to the EWTcanopy values where LAI is equal to 1. As mentioned above, soil effects cause the regression line to
Fig. 4. Sensitivity of GVMI to EWTcanopy with respect to various biophysical factors, geometry of observation, and atmospheric conditions. LAI 1 indicates those regions in which the LAI = 1. These regions show greater dispersion due to soil effects.
196
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197
shift towards the soil reflectance values. The GVMI is therefore suitable for retrieving vegetation water content when the LAI is equal to or greater than 2. For sparsely vegetated area, where LAI is less than 2, further research is required to understand the role of soil effects on reflectance measured in the all-optical spectrum. Further research is therefore required in order to propose an index that allows the retrieval of EWTcanopy for sparsely vegetated area. 3.3. Performance of the index Variability of the index values with respect to changes in EWTcanopy is illustrated in Fig. 4. The variability is obtained using simulations performed with all scenarios described in Section 3.1. The relatively low dispersion of these values around the best-fitted line indicates that the index provides a reliable estimator of the quantity of water per unit area at canopy level. The performance of GVMI in retrieving EWTcanopy values is measured by a signal-to-noise (S/N) ratio and the root mean square (RMS) as suggested by Leprieur, Verstraete, and Pinty (1994) and applied by Gobron et al. (2000). Between 0 and 2100 g m 2, the RMS equals 0.046 and the S/N equals 11.28 indicating that the GVMI shows good sensitivity to the EWTcanopy rather than to the noise. This S/N value is higher than the S/N value obtained for NDVI (6.22) using SPOT-VEGETATION wavelengths (Gobron et al., 2000). However, between 2100 and 3500 g m 2, the RMS equals 0.02 and the S/N drops to 1.68 indicating that, for high values of EWTcanopy, variation in EWTcanopy values is not reflected by a GVMI variation. There is a phenomenon of saturation that will prevent the detection of EWTcanopy variations in certain ecosystems where the quantity of water is greater than 2100 g m 2. This saturation phenomenon was also reported for high EWT values at leaf level by Ceccato et al. (2001). To validate the index, some measurements were performed in different ecosystems ranging from a shrub steppe to a savannah woodland in Senegal. Results are published and discussed in a separate article (‘‘Designing a Spectral Index to Estimate Vegetation Water Content from Remote Sensing Data: Part 2. Validation and Applications’’).
4. Conclusion Using a GSA, we have quantified the importance of vegetation water content and other factors on reflectance measured at three levels (leaf, canopy, and atmosphere). We have demonstrated that variations in vegetation water content significantly affect the reflectance measured in the SWIR. We have also demonstrated that a combination of the SWIR and the NIR is required to retrieve vegetation water content. Using this information, an index was created that optimises the retrieval of vegetation water content expressed in terms of quantity of water in the canopy per
unit area (EWTcanopy) and minimises the perturbing effects of geophysical and atmospheric effects. The index GVMI was specially optimised for the new SPOT-VEGETATION sensor as an example of how the method could be applied. Performances of GVMI were also discussed. Validation of the GVMI, compared with field measurements in different ecosystems, is presented in a separate article (‘‘Designing a Spectral Index to Estimate Vegetation Water Content from Remote Sensing Data: Part 2. Validation and Applications’’).
Acknowledgements This work was supported by the D.G. Joint Research Centre of the European Commission under a Marie Curie Research Training Grant to the first author.
References Asner, G. P. (1998). Biophysical and biochemical sources of variability in canopy reflectance. Remote Sensing of Environment, 64, 234 – 253. Asner, G. P., Braswell, B. H., Schimel, D. S., & Wessman, C. A. (1998). Ecological research needs from multiangle remote sensing data. Remote Sensing of Environment, 63, 155 – 165. Bacour, C., Jacquemoud, S., Tourbier, Y., Dechambre, M., & Frangi, J.-P. (2002). Design and analysis of numerical experiments to compare four canopy reflectance models. Remote Sensing of Environment, 79, 72 – 83. Campolongo, F., Saltelli, A., Sørensen, T., & Tarantola, S. (2000). Hitchhiker’s guide to sensitivity analysis. In A. Saltelli, K. Chan, & E. M. Scott (Eds.), Sensitivity analysis ( pp. 15 – 47). New York: Wiley. Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., & Gre´goire, J.-M. (2001). Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sensing of Environment, 77, 22 – 23. Cukier, R. I., Fortuin, C. M., Schuler, K. E., Petschek, A. G., & Schaibly, J. H. (1973). Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients: I. Theory. Journal of Chemical Physics, 59, 3873 – 3878. Cukier, R. I., Levine, H. B., & Schuler, K. E. (1978). Nonlinear sensitivity analysis of multiparameter model systems. Journal of Computational Physics, 26, 1 – 42. Danson, F. M., Steven, M. D., Malthus, T. J., & Clark, J. A. (1992). Highspectral resolution data for determining leaf water content. International Journal of Remote Sensing, 13(3), 461 – 470. Dawson, T. P., Curran, P. J., North, P. R. J., & Plummer, S. E. (1999). The propagation of foliar biochemical absorption features in forest canopy reflectance: a theoretical analysis. Remote Sensing of Environment, 67, 147 – 159. Diner, D., Asner, G., Davies, R., Knyazikhin, Y., Muller, J.-P., Nolin, A., Pinty, B., Schaaf, C., & Stroeve, J. (1999). New directions in earth observing: scientific applications of multiangle remote sensing. Bulletin of the American Meteorological Society, 80, 2209 – 2228. Downing, H. G., Carter, G. A., Holladay, K. W., & Cibula, W. G. (1993). The radiative-equivalent water thickness of leaves. Remote Sensing of Environment, 46, 103 – 107. Dull, C. W., & Singh, A. (2000). Satellite earth observation for wildland fire management. In F. Ahern, J.-M. Gre´goire, & C. Justice (Eds.), Forest fire monitoring and mapping: a component of Global Observation of Forest Cover. Report of a workshop, Joint Research Centre, Ispra, Italy, November 3 – 5, 1999 ( pp. 85 – 97). Italy: European Commission, Joint Research, EUR 19588 EN.
P. Ceccato et al. / Remote Sensing of Environment 82 (2002) 188–197 Gao, B.-C. (1996). NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 58, 257 – 266. Gastellu-Etchegorry, J. P., Guillevic, P., Zagolski, F., Demarez, V., Trichon, V., Deering, D., & Leroy, M. (1999). Modeling BRF and radiation regime of boreal and tropical forests: I. BRF. Remote Sensing of Environment, 68, 281 – 316. Gobron, N., Pinty, B., Verstraete, M., & Govaerts, Y. (1999). The MERIS Global Vegetation Index (MGVI): description and preliminary application. International Journal of Remote Sensing, 20(9), 1917 – 1927. Gobron, N., Pinty, B., Verstraete, M. M., & Govaerts, Y. (1997). A semidiscrete model for the scattering of light by vegetation. Journal of Geophysical Research, 102, 9431 – 9446. Gobron, N., Pinty, B., Verstraete, M. M., & Widlowski, J.-L. (2000). Advanced vegetation indices optimized for up-coming sensors: design, performance and applications. IEEE Transactions on Geoscience and Remote Sensing, 38(6), 2489 – 2505. Govaerts, Y. M., Verstraete, M. M., Pinty, B., & Gobron, N. (1999). Designing optimal spectral indices: a feasibility and proof of concept study. International Journal of Remote Sensing, 20(9), 1853 – 1873. Hunt, Jr., E. R., & Rock, B. N. (1989). Detection of changes in leaf water content using near- and middle-infrared reflectances. Remote Sensing of Environment, 30, 43 – 54. Jacquemoud, S. (1993). Inversion of the PROSPECT + SAIL canopy model from AVIRIS equivalent spectra: theoretical study. Remote Sensing of Environment, 44, 281 – 292. Jacquemoud, S., Bacour, C., Poilve´, H., & Frangi, J.-P. (2000). Comparison of four radiative transfer models to simulate plant canopies reflectance—direct and inverse mode. Remote Sensing of Environment, 74 (3), 471 – 481. Kaufman, Y. J. (1989). The atmospheric effect on remote sensing and its correction. In G. Asrar (Ed.), Theory and applications of optical remote sensing ( pp. 336 – 428). New York: Wiley. Kaufman, Y. J., & Tanre´, D. (1992). Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Transactions on Geoscience and Remote Sensing, 30(2), 261 – 270. Leprieur, C., Verstraete, M. M., & Pinty, B. (1994). Evaluation of the
197
performance of various vegetation indices to retrieve vegetation cover from AVHRR data. Remote Sensing Reviews, 10, 265 – 284. Lillesaeter, O. (1982). Spectral reflectance of partly transmitting leaves: laboratory measurements and mathematical modeling. Remote Sensing of Environment, 12, 247 – 254. Mayaux, P., Gond, V., & Bartholome´, E. (2000). A near-real time forestcover map of Madagascar derived from SPOT-4 VEGETATION data. International Journal of Remote Sensing, 21(16), 3139 – 3144. Pen˜uelas, J., Filella, I., Biel, C., Serrano, L., & Save´, R. (1993). The reflectance at the 950 – 970 nm region as an indicator of plant water status. International Journal of Remote Sensing, 14(10), 1887 – 1905. Price, J. C. (1995). Examples of high resolution visible to near-infrared spectra and a standardized collection for remote sensing studies. International Journal of Remote Sensing, 16(6), 993 – 1000. Roberts, D. A., Green, R. O., & Adams, J. B. (1997). Temporal and spatial patterns in vegetation and atmospheric properties from AVIRIS. Remote Sensing of Environment, 62, 223 – 240. Saltelli, A. (2000). What is sensitivity analysis? In A. Saltelli, K. Chan, & E. M. Scott (Eds.), Sensitivity analysis ( pp. 3 – 13). New York: Wiley. Saltelli, A., Tarantola, S., & Chan, K. P.-S. (1999). A quantitative modelindependent method for global sensitivity analysis of model output. American Statistical Association and the American Society for Quality. Technometrics, 41(1), 39 – 56. Tucker, C. J. (1980). Remote sensing of leaf water content in the near infrared. Remote Sensing of Environment, 10, 23 – 32. Ustin, S. L., Roberts, D. A., Pinzon, J., Jacquemoud, S., Gardner, M., Scheer, G., Castan˜eda, C. M., & Palacios-Orueta, A. (1998). Estimating canopy water content of chaparral shrubs using optical methods. Remote Sensing of Environment, 65, 280 – 291. Vermote, E. F., Tanre´, D., Deuze´, J. L., Herman, M., & Morcrette, J.-J. (1997). Second simulation of the satellite signal in the solar spectrum, 6S: an overview. IEEE Transactions on Geoscience and Remote Sensing, 35(3), 675 – 686. Verstraete, M. M., & Pinty, B. (1996). Designing optimal spectral indexes for remote sensing applications. IEEE Transactions on Geoscience and Remote Sensing, 34(5), 1254 – 1265.
