Note on the compatibility of ICOS, NEON, and TERN sampling designs, different camera setups for effective plant area index estimation with digital hemispherical photography

Leaf Area Index (LAI), a measure of the amount of plant leaf material in an ecosystem, is defined as one half of the total green leaf area per unit ground surface area (Chen & Black, 1992; Fernandes et al., 2014). LAI is included in many models describing vegetation-atmosphere interactions (GCOS-138, 2010) as a critical variable controlling processes such as photosynthesis, respiration, and rain interception. LAI estimates using different Earth Observation (EO) data and algorithms have been developed and made available to the user community over the last decades (Fang et al., 2019) as ready-for-use products. When no adjustments are made for the clumping effect or for the interception of light by woody canopy elements when estimating LAI, the variable is the effective Plant Area Index (PAI_e) instead (Chen, 1996).

The Global Climate Observing System (GCOS) has specified the need to update and validate global LAI predictions systematically. The Land Product Validation (LPV) sub-group of the Committee for Earth Observation Satellites (CEOS) developed a strategy for validating the predictions (Fernandes et al., 2014). The CEOS LPV currently encourages validation and intercomparison of the so-called LPV Supersites. The majority of the LPV Supersites come from well-known and established networks, including the Integrated Carbon Observation System (ICOS) (Gielen et al., 2017) in Europe, the National Ecological Observatory Network (NEON) (Meier & Jones, 2018) in the U.S., and the Terrestrial Ecosystem Research Network (TERN) (Karan, 2015) in Australia. However, the networks employ different sampling designs (i.e. number of measurements, location of sampling points within the test site) and camera setups for in situ LAI estimation. Previous studies have reported apparent differences in the LAI estimates from different sampling schemes (Calders et al., 2018; Liu et al., 2021; Majasalmi et al., 2012; Nackaerts et al., 2000; Zou et al., 2020).

The objective of this study is to compare the three different schemes adopted within the ICOS, NEON, and TERN networks instead of searching for the most optimal, new sampling scheme. We carried out the intercomparison measurements with digital hemispherical photography during diffuse light conditions, adopting the established ICOS, NEON, and TERN sampling schemes. We use data from a measurement campaign conducted in 2021 at the Järvselja Radiation Transfer Model Intercomparison (RAMI) (Widlowski et al., 2015) birch stand. The impacts of the sampling schemes and camera setups on the gap fraction and (PAI_e) prediction and estimation from digital hemispherical photography (DHP) were analysed.

The layout of image locations, i.e. sampling grid has some effect on the gap fraction estimate at stand level (Figure 2). The most significant differences between the sampling schemes occur near the zenith and smaller viewing angles. The differences in camera setups can also be noticed. Results from all setups correlate well with each other with R² values around 0.99, but setups I and II give a more similar gap fraction across view angles (average difference 0.005), whereas setup III deviates more from setups I and II (average difference 0.01 from setup I). Gap fractions from setups I and II consistently differ a little. Setup I almost always has a slightly smaller value for the gap fraction in all viewing zenith angles except the very small and the very large ones (Figure 2).

Figure 2.

Based on the gap fraction, PAI_e estimates also vary between schemes and camera setups. The overall PAI_e, for the test site based on all schemes and all camera setups, comprising a total of 324 images, was 3.09 with a standard error (SE) of 0.02. Following Majasalmi et al. (2012), it is assumed that this very large set of 324 measurements is representative of the population mean. Here it is assumed that there is no influence of different measurement devices and operators on observations, however, later we show that each measurement series with a particular device had its own uncertainty characteristics. The ICOS and TERN schemes result in slightly smaller estimations for the PAI_e (2.99 and 3.03, respectively) than the NEON (3.26) scheme (Table 2). As the gap fraction estimation was consistently different between setups I and II, the resulting PAI_e estimation also reflects this difference, and the PAI_e values are consistently a little bit greater for setup I than for setup II. The PAI_e estimation for setup III stays close to the estimates of other two setups. Overall average PAI_e estimations for all setups and schemes ranged from 2.90 to 3.38, with standard errors ranging from 0.04 to 0.06 (Table 2).

All three cameras took images at the same marked locations within a few minutes. Still, there are always slight deviations in positioning the cameras precisely at the same spot with identical levelling and orientation. This causes some variability in the recording of the scene content. Each camera operator processed the images taken with a particular camera. This also introduces some small variability into the results because (1) the markers of open sky pixel sampling markers can be freely chosen and (2) identification of canopy gaps at larger view zenith angles for extracting sky pixel samples requires some practicing and experience. However, Lang et al. (2013) show that the influence of randomness in open sky sampling point locations due to different analysts is marginal for the LinearRatio_SC. The PAI_e calculated from camera setups I and II has a very strong correlation (R² = 0.94) within image pairs, but setup III varied more in the results, giving often quite large deviations in the sampling points compared to setups I and II (Figure 3). Although setups I and II correlate well, the linear model fitted on the data has a slope around 0.91 and is also slightly shifted away from the 1 to 1 line, indicating a small systematic difference in the PAI_e.

Figure 3.

As the forest stand is natural, i.e. not perfectly homogeneous, it can be expected that the PAI_e obtained at different measurement points would vary, for example for setup I, the PAI_e calculated from single images ranges from 2.45 to 4.06 (Figures 4, 5, 6). However, the PAI_e includes both – the true value and uncertainty that is introduced during the measurements and data processing.

Figure 4.

Figure 5.

Figure 6.

Figure 7.

Setup III had a problem with overexposure due to the semiautomatic imaging mode. Some images were out of focus – possibly due to accidental turning of the focus wheel, resulting in loss of detail in smaller gaps, leaves and branches (Figure 8). Even though setup III had the highest pixel resolution, the loss of small details due to the wrong focus setting combined with overexposure did contribute to inconsistent results at many sampling points demonstrated by PAI_e values compared to setups I and II (Figure 3B, C). We identified overexposed measurements of setup III based on possible maximum pixel values according to the camera radiometric resolution (also known as bit depth) and found that the concordance of the rest (i.e. correct measurements with setup III) with the other measurements was strong (Figure 3B, C).

Figure 8.

The impact of image overexposure on the measurements of the canopy gap fraction and consequently PAI_e was not constant. The PAI_e of some of the overexposed measurements was not much different from setup I and setup II, which can be attributed to the magnitude of overexposure and also to the amount of mixed pixels in the scene where “burn-in” effects occur first. Additionally, the out-out-of-focus problem tended to fill small canopy gaps of few pixels in size, effectively decreasing the canopy gap fraction. Therefore, it is the responsibility of the camera operator to keep the camera in the linear sensitive range of the sensor to avoid overexposure and systematic errors in the canopy gap fraction and PAI_e.

The PAI_e values obtained at each sampling point correlated well between setups I and II with R² values of 0.97 for ICOS, 0.89 for NEON, and 0.92 for TERN (Table 3). The differences in correlation are probably due to the uncertainty of measurements and cannot be attributed to a particular sampling scheme. Overexposed measurements decreased substantially the correlation of setup III with the others. However, in the subset of correct measurements that did not suffer from overexposure, the correlations were of the same strength than between setup I and setup II (Table 3).

As we noticed some systematic differences in camera setups I and II, we conducted an experiment where the images of the ICOS scheme were cross-processed by the camera operators to see if there was some human error or bias. We found a small difference in the estimated gap fraction around the large viewing angles (Figure 8) but overall, the results are not statistically different from each other with a one-tail t-test p-value of 0.486.

An estimation of plant or leaf area index for a test site using optical measurements of canopy gap fraction or gap size distribution is influenced by many factors that introduce uncertainty into the results. In our study, we focused on the sampling design, field measurements, and image processing. Sampling design, in an ideal case, (1) provides that enough measurements are made to obtain sufficiently narrow confidence intervals for the estimates and (2) ensures that repeated measurements in time series are comparable by fixing, for example, the arrangement of permanent and random sampling points. The stage of field measurements involves uncertainties related to configuring cameras, locating measurement points, setting up the camera on the point, observing illumination conditions according to the measurement protocol, determining exposure settings to keep the camera in its linear sensitivity range, and the physical process of taking images. Image processing involves extracting the data from camera output files, applying radiometric and geometric corrections, selecting the image processing method and its parameters and software, and calculating the gap fraction according to the view zenith angle from the processed images for the Equation (2). Prediction of the true green leaf area index further from the canopy gap fraction includes another wide range of models and methods.

We used three different sampling schemes and obtained the PAI_e in the range of 2.99–3.26 for the Järvselja RAMI birch stand. The result is close to Kuusk et al. (2009), who reported PAI_e 2.94 measured at nine positions with the LAI-2000 instrument within the same test site. The small increase in PAI_e values compared to Kuusk et al. (2009) can be explained by a growth of dominant and lower tree layers during the period (>15 years) between the measurements (Lang & Pisek, 2019). The PAI_e derived from ICOS and TERN sampling schemes were more similar to each other (two-tailed t-test p-value 0.229) than NEON (t-test p-value < 0.00001) results (Table 2). The difference between NEON results compared to the other schemes may be explained by the distribution of the sampling points in the forest. The NEON scheme has three tightly packed locations in the forest where images are taken, whereas TERN and ICOS are more spread out over the whole plot (Figure 1). With NEON, the initial selection of the three sub-section locations is important because if two or all three locations fall into either more open or closed canopy areas, the sample is not representative of the forest as a whole. Even though we followed the scheme layout instructions carefully, with our testing, two of the locations happened to have rather closed canopy conditions and the resulting images gave on average a smaller gap fraction when compared to the ICOS and TERN sampling schemes (Figure 2). The NEON protocol uses this kind of sampling in conjunction with a second larger set of distributed plots that are sampled less frequently but cover more of the forests. The tower plots described in the NEON protocol can give a good time series of the area near the sampling points, but they should be viewed together with the distributed plots results to properly represent a forest site.

Proper camera configuration is essential to get consistent and comparable measurements. As our goal was to measure incident radiation as precisely as possible, it was necessary to avoid over- or underexposing the image. In our testing, setup III illustrated possible issues with overexposed and out-of-focus images where smaller details of the canopy were often lost (Figure 7C). This resulted in occasional inconsistent results when compared to setups I and II, which had a very good correlation with each other (Figure 3A). When tested afterwards, reducing the focus setting of the camera decreased the effective PAI_e by 1.3% on average when the focus was turned down from infinity to 0.2 meters. When intentionally overexposing the images the PAI_e decreased significantly. When an image with no overexposed pixels was compared to another one that had 0.1% of pixels overexposed to the maximum digital value of the sensor, the PAI_e decreased by around 1.7%. When the amount of overexposed pixels was around 0.4%, the PAI_e decreased substantially by 14%. Here we see that for obtaining data for the canopy gap fraction it is crucial to operate hemispherical cameras in their linear sensitive range.

Our intercomparison test allowed us to search for possible influences of cameras and image analysts on the canopy gap fraction and calculated PAI_e. There was very strong correlation between the PAI_e of setups I and II with a small systematic difference. In a time series this could be interpreted as a possible change in the PAI_e, but in our experiment the measurements were done almost within a few minutes in time that excludes the change in the PAI_e. For image processing we used a well-established method LinearRatio_SC that, by its principle, is an analogue for LAI-2000 when compared to various threshold-based and pixel classification-based algorithms. To apply LinearRatio_SC the image analyst has to place sky pixel sampling marks for constructing open sky reference which involves a subjective decision to some extent.

We tested the influence of operator subjectivity by cross-reprocessing the ICOS scheme images that were taken with setup I and II by corresponding operators. However, no statistically relevant differences were found in the gap fraction calculated from the same set of images processed by two different people (Figure 8). This indicates that there is a more fundamental reason stemming from the raw images or cameras themselves. We also tested the effects of vignetting and projection models, subtracting the dark images, and the effects of different versions of dcraw. They all had negligible or no impact on the results. Setup I consistently had a little but smaller gap fraction than setup II. Addressing all these possible sources of uncertainty, we conclude that the difference may be caused by light scattering and colour crosstalk within cameras. Another possible reason could be a small systematic difference in camera height over the ground or vertical levelling during field measurements. For example, a two-degree error in camera levelling changes the path length at 57° view zenith angle up to three metres (10%) in the stand. As a conclusion, we recommend always to carry out intercomparison measurements with old and new cameras when devices are upgraded.

Overall, we illustrate that the three different employed schemes produce comparable results for estimating the gap fraction and PAI_e. Although different setups and schemes gave differing results, with Anova Two-Factor test p-values <0.001, there was no connection between the camera setup used and the scheme (p-value 0.178). Our testing shows that more evenly spread out sampling schemes are more consistent with each other, exemplified by ICOS and TERN, and could be more representative of the overall conditions of the forest. The results vary when different cameras with different operators are used, but it appears that 36 images employed by all three setups are enough to deliver an average with sufficiently narrow confidence limits. It depends on the application what type of cameras and how high accuracy of calibration is required. From this study we cannot conclude which scheme or camera setup is best used in the field, because all these established schemes vary slightly in their intended use and purpose and are not necessarily intended to yield site level mean estimates. But we did conclude that some layout schemes may have inherited problems when applied to estimate the PAI_e for our 1-hectare plot. The CEOS Land Product Validation Subgroup (Morisette et al., 2006) suggests in their proposed framework, for validating LAI products, to use a two-staged sampling design. In the case of NEON and ICOS, the circular and cross shaped compact sampling areas would be considered elementary sampling units used together with high spatial resolution satellite data to capture the variability in the site extent. These two sampling schemes are not directly designed to capture the average PAI_e for a 1-hectare site as in our case.

The CEOS LPV is currently in the process of selecting supersites with a fully characterized land surface and vegetation cover for the parameterization of 3D radiative transfer models. Our study contributes towards establishing the procedures for uncertainty estimation by evaluating potential systematic errors stemming from collecting in situ measurements with DHP using different scheme designs, methods and devices currently in use at most LPV Supersites.

Future work can include efforts to investigate the impact of sampling schemes on the LAI estimation of forests from other indirect LAI measurement methods (e.g. terrestrial laser scanner and LAI-2200). Ground-Based Observations for Validation (GBOV) of Copernicus Global Land Products (GBOV, 2023) uses schemes studied in this paper to validate its decametric and hectometric satellite-derived LAI products (Brown et al., 2020) and thus it is important to study these schemes. More comparisons between measurement methods and data processing methods would be useful in determining the most reliable ways to gather data on the gap fraction. Forest plots with different plant functional types may also be included to consolidate the findings of this study.