Previous PageTable Of ContentsNext Page

Interpolating leaf area index between destructive sampling dates using remote sensing in the Australian Grains Free Air Carbon dioxide Enrichment (AGFACE) experiment

Glenn Fitzgerald1, Rob Norton2, Garry O’Leary1 and Mahabubur Mollah1

1 Department of Primary Industries, 110 Natimuk Rd., Horsham, VIC 3401. Email
University of Melbourne, 110 Natimuk Road, Private Bag 260, Horsham, Victoria, Australia


The Australian Grains Free Air Carbon dioxide Enrichment (AGFACE) experiment in Horsham, Victoria, was designed to simulate the elevated atmospheric carbon dioxide (CO2) concentrations expected to occur in 2050, in part to provide data for validation of crop models. The experiment measures interacting effects of different atmospheric CO2 concentrations, irrigation, time of sowing, nitrogen fertiliser, and cultivar on wheat growth and production. Carbon dioxide was injected over the crop in open-air 12 m rings. Destructive samples were collected at stem elongation (DC30), anthesis (DC65) and harvest (DC92). Important crop attributes, such as leaf area index (LAI) and biomass were collected at these sampling times but, due to limited space, more destructive measurements to track seasonal trends could not be sampled. This study explored the use of the remote sensing Normalised Difference Vegetation Index (NDVI) as a surrogate to estimate LAI at dates between destructive sampling times. The relationship predicting LAI from NDVI, using treatment mean data from both DC30 and DC65 had r2 = 0.79 and root mean square error = 0.22 LAI (m2/m2). Also, the pattern of statistical significance for differences among treatments was similar at both DC30 and DC65 for both LAI and NDVI. These preliminary findings suggest that NDVI may be used as a surrogate to estimate LAI at dates between destructive sampling dates.

Key Words

FACE, LAI, remote sensing, NDVI, interpolation


Free Air Carbon dioxide Enrichment (FACE) experiments were developed to elevate CO2 levels around plants in as realistic a way as possible, to study the effects on plant growth and development. Open rings are used to distribute CO2 over the crop to simulate future CO2 levels. The AGFACE experiment was established to quantify the response of wheat to elevated CO2 to better understand the crop response to climate change.

Providing data for modeling the growth and development of wheat canopies under future climate change scenarios is one of the main objectives of the AGFACE experiment. Since destructive sampling occurs only at the developmental stages (decimal code) DC30, DC65 and DC92 developmental stages, effects of elevated CO2 and interactions with other treatments at other times cannot be assessed directly. Remote sensing can provide an indirect measure to understand canopy development between the sample collection dates.

Leaf area index (Leaf area/ground area, LAI) is an important input variable to crop models for the AGFACE project. The Normalised Difference Vegetation Index (NDVI) has been used for many years to estimate LAI, biomass and other parameters of crop canopies (Carlson and Ripley, 1997). Once a relationship between the LAI and NDVI from the destructive samples is established, the LAI can be interpolated from NDVI at dates where LAI has not been measured. These data can then be used to validate the model output and, if necessary, provide information to update or correct model parameters. A preliminary assessment of the NDVI and LAI relationship was undertaken in this study.


The FACE experiment was established at the Department of Primary Industries Plant Breeding Centre in Horsham, Victoria, in 2007. The site was planted to wheat (Triticum aestivum L.) with 24 treatment combinations, constituted from two levels of CO2 (ambient aCO2, elevated eCO2 – 550 ppm), two levels of irrigation (Irrigated (I), Rainfed (R)), two times of sowing (early, TOS1 and late, TOS2), and three variety-nitrogen treatments (Yitpi-N0, Yitpi-N+, Janz-N0). Each ring was either I or R. The four CO2-irrigation treatments were applied to 16 rings (Main Plots) arranged in a 4x4 row-column layout similar to a Latin square. Each ring (octal) had a diameter of 12 m and was located in the centre of a 20 X 20 m (400 m2) area planted to wheat. Each ring was divided into two half rings which received the two TOS treatments. Each half ring was split into six sub-plots, each of size 1.6 x 4.0 m, three designated for intra-season growth measurements and three for measurements at maturity. The three variety-nitrogen treatments were randomly allocated to the three growth stage and the three maturity sub-plots (Figure 1). Destructive samples were taken from within the growth stage sub-plots at DC30 and DC65 for each TOS. Final harvest material was collected from the maturity sub-plots.

Figure 1. Example of an AGFACE ring and sub-plot layout. Each ring was assigned a CO2 X Irrigation treatment (ie, aCO2 x I, aCO2 x R, eCO2 x I, eCO2 xX R). Letters designate each of the sub-plots.

Plant samples for LAI were collected from the growth sub-plots at both DC30 and DC65 (4 X 0.5 m of drill rows). The NDVI data were collected from both growth and maturity sub-plots at 14 dates throughout the season, including DC30 and DC65. The NDVI data from the growth and the maturity plots were respectively compared to the DC30 and the DC65 LAI data. This was done because after DC30 the growth plots had gaps that changed the relationship of LAI to NDVI. Although DC65 biophysical data were not collected from the maturity plots, these represented undisturbed areas, more similar to the areas within the growth plots at DC30.

Remote sensing data were collected using an active sensor (Crop Circle, Holland Scientific, USA). This sensor collects two wavebands of information in the Red and NIR portions of the spectrum and outputs an NDVI value. It has a built-in light source so is an “active” sensor. The sensor was mounted on a handheld boom and “walked” across the plots approximately fortnightly throughout the season. The instrument collects ten readings per second with 30-50 acquired from each sub-plot. The NDVI were averaged at sub-plot level for use in further analysis.

We used two criteria to investigate the appropriateness of using NDVI as a possible surrogate for LAI. The first was to assess the nature and the strength of relationship between LAI and NDVI through regression analysis. The second criterion was to assess if LAI and NDVI, at both DC30 and DC65, delivered a similar pattern of significance of differences among treatments. For this, the sub-plot level data on LAI and NDVI were analysed in GenStat software as per experimental design using either the analysis of variance (ANOVA) at DC30 or the restricted maximum likelihood (ReML) at DC65 due to outliers or missing data (2 out of 48 observations excluded). The NDVI data were scaled by setting soil values to 0 and the seasonal maximum to 1 (Carlson and Ripley, 1997). Data more than 3 standard deviations from the treatment means were removed from analysis.


There were four data sets available for analysis: TOS1 and TOS2 at both DC30 and DC65. TOS 1 and 2 were sown nine weeks apart (Table 1) to provide differences in yield and different temperature regimes during grain fill.

Table 1. Agronomic dates, FACE 2007


Sowing date

Destructive sample dates


TOS 1, DC30

18 Jun 07

06 Sep 07


TOS 1, DC65

18 Jun 07

29 Oct 07


TOS 2, DC30

23 Aug 07

25 Oct 07


TOS 2, DC65

23 Aug 07

20 Nov and 3 Dec 071


1TOS-time of sowing.2 Part of TOS2 was not at DC65 so sampled when reached this stage.

Analysis of variance at DC30 and DC65 for TOS1 (Table 2) showed that at DC30 there were no significant effects of the main treatments on biomass, LAI or NDVI. The one interaction of CO2 x variety or N was significant for biomass and LAI at α=0.1 but the NDVI did not detect this. At DC65, there were several

significant treatment effects (p<0.05) with NDVI and LAI showing similar treatment significance pattern. The ANOVA results thus seem to exhibit a similar pattern of treatment significance at both DC30 and DC65 for LAI as well as NDVI. To give a sense of the actual treatment differences, at DC65 for example, the range of biomass across treatments was 1.72 (aCO2, R) to 2.01 (eCO2, I) t/ha and LAI ranged from 1.19 (eCO2, R) to 1.75 (eCO2, I).

Table 2. Effects of treatments on aboveground biomass, LAI and NDVI at TOS1, ANOVA, 2007 (significant relationships only, p ≥ 0.1).


Aboveground biomass



-------------------- DC30, 6 Sep 07 (p values) ----------------------





CO2 X (V or N)




Irrigation (Irr)




Var or N




--------------------- DC65, 29 Oct 07 (p values) ---------------------









I X (V or N)




V or N




1 NDVI at DC30 collected from “Growth” plots, which were destructively sampled at DC30 and DC65.

2 NDVI at DC65 collected from “Maturity” plots, which were destructively sampled at harvest.

The treatment-means-based regression (Fig. 2) of LAI on NDVI had slope = 3.98 and intercept = -1.26 (both had p < 0.0001), with RMSE = 0.22 LAI (m2/m2) and r2 = 0.79. This was based on n=45 (3 points removed > 3 s.d. from mean) treatment means at TOS1-DC30, TOS1-DC65 and TOS2-DC65 corresponding to the four CO2-irrigation and two variety-nitrogen treatments (Yitpi-N0, Yitpi-N+). This did not include TOS2-DC30 data as this sampling showed unusually high LAI relative to NDVI. The reasons for this have yet to be investigated. Regression residuals were approximately normally distributed and showed uniform scatter.

Figure 2. Plot and treatment mean NDVI scaled to soil and maximum canopy values vs. Leaf area index (LAI) for sub-plot data (〉, thin line) and treatment means (●, thick line).

Using the sub-plot level data (n=138), the regression of LAI on NDVI had slope = 2.93 and intercept = -0.69, (both had p < 0.0001) with RMSE = 0.39 m2/m2 and r2 = 0.47. Low values of NDVI (below 0.3) are not reliable measures of plant biophysical features (Baret and Guyot, 1991) so four points were excluded from the plot level data and one from the mean data set (which was also an outlier). Two other outliers were removed from the mean data (residuals more than 3 s.d. from mean).

Much of the variation was likely caused by different spatial sampling scales of the biophysical and remote sensing data. The NDVI data represented sub-plot means. The LAI data were obtained from the biophysical samples and represented 2 m of drill row within each sub-plots. All sensors are sensitive to the amount of living vegetation in their field of view. Plant responses that affect the amount of vegetation visible from above the canopy will change sensor values. For example, water stress can cause leaf angle changes (drooping and wilting) that cause a reduction in NDVI values while not affecting the physical measure of LAI. A scaled NDVI was used to normalise the data to the maximum and minimum range of NDVI values for soil and maximum canopy (Carlson and Ripley, 1997) resulting in values of 0 to 1 that represent the range of NDVI present during the season. This can potentially provide a more consistent measure of canopy cover (or LAI) than the NDVI alone when comparing across locations or dates.

It is well established that before full cover is reached (about LAI = 3) there is a linear or near-linear relationship between LAI and NDVI (Carlson and Ripley, 1997). In rainfed wheat in Victoria, LAI values as high as 3.0 are uncommon. Thus, NDVI vs. LAI relationships should have minimal non-linear features (Fig. 2). However, as pointed out by Carlson and Ripley (1997), LAI, cover, and NDVI are not independent. Estimation of LAI from NDVI depends on fractional cover, which is a function of LAI. In the rainfed wheat systems of Victoria, a direct measure of LAI from NDVI may require adjustment for fractional cover. This may be the reason the intercept in Figure 2 does not pass through zero. This will be investigated in future analyses.

A time series of NDVI data shows how the eight treatments for the Yitpi wheat variety differed across the season (Fig. 3). There are few differences at DC30, evident also in the ANOVA data (Table 2). Later, canopy responses diverged with some “crossing over” or treatments having greater relative NDVI values early in the season and lower relative values later. This type of temporal data can help validate crop models.

Figure 3. Time series NDVI for TOS1, Yitpi variety, Maturity plots. I = irrigated, R= rainfed, N0 = no added nitrogen fertiliser, N+ = added N fertiliser, aCO2 = ambient CO2 concentration and eCO2 = elevated CO2 concentration.


The objective of using remote sensing data (NDVI) as a surrogate for LAI is to “fill in” the gaps between destructive sample dates in this FACE experiment is to understand temporal patterns and as a validation for crop models. The relationship between NDVI and LAI is established at the destructive sampling times. Analysis of variance of the sub plot data showed that the NDVI and LAI elicited similar significance patterns. Regression modelling of treatment means using NDVI to predict LAI established the relationship to be near linear with a acceptable correlation coefficient. Thus NDVI could be used to track changing LAI across the season and for validation of crop models based on treatments means. It is unclear whether sub-plot level inferences can be made due to the high variance in these data. The high variance was likely due to sampling methods that will be modified in the following seasons. Future analyses will include more years of data as well as testing other remote indices that can be derived from the two wavebands available and more sophisticated analytical techniques.


Baret F and Guyot G (1991). Potential and limits of vegetation indices for LAI and APAR assessment. Remote Sensing and Environment 35:161-173.

Carlson TN and Ripley DA (1997). On the relation between NDVI, fractional vegetation cover and leaf area index. Remote Sensing and Environment 62:241-252.

Previous PageTop Of PageNext Page