Introduction:

The boreal forest represents 11% of the earth’s total land area ( Bonan and Shugart, 1989), and contains more than 30% of all forest soil carbon (Schlesinger, 1991). Global climate simulations indicate that the boreal region will undergo significant warming in response to increasing atmospheric CO₂ (Houghton et al., 1990). Due to the large contribution of these forests to the global carbon budget, warming in these northern regions could have large implications for the global carbon cycle and climate regulation.

The need to better understand how the boreal forest will respond to climate change has been the main driving force behind a large scale study of the Canadian boreal region, named the "Boreal Ecosystem-Atmosphere Study" (BOREAS, Sellers et. al, 1995). One need of this study was to improve our current understanding of boreal carbon fluxes. Recently, there has been an ongoing debate about whether the boreal forest is a net sink or source of terrestrial carbon (Wang and Polglase, 1995). Some studies have suggested that it is a source (Quay et al., 1992). Other studies have suggested the opposite, that it is a sink, thus helping to offset the rising atmospheric CO₂ concentration (Keeling et al, 1996). A major focus of the BOREAS project was to examine the exchange of CO₂ between the boreal ecosystems and the atmosphere to clarify the spatial and temporal patterns of carbon flux.

Remote sensing provides one method of exploring spatial patterns of carbon flux. Multiple studies have shown that the optically based Normalized Difference Vegetation Index (NDVI) can be used to estimate carbon fluxes of vegetation (Goward et. Al., 1985, Whiting et al., 1991). NDVI is a function of the fraction of photosynthetically active radiation (f_PAR) absorbed by green vegetation (or f_APAR). When combined with a measurement of the photon flux density of photosynthetically active radiation (PFD_PAR), f_APAR can be used to determine the amount of photosynthetically active radiation absorbed by green vegetation (APAR):

APAR = f_APAR x PFD_PAR (1)

APAR can be multiplied by an efficiency factor (?) to derive the net primary productivity (NPP) or the net photosynthetic carbon gain of a biome type:

Photosynthetic rate = (? x APAR) (2)

Usually, efficiency (?) is not derived directly from remote sensing. However, some recent studies have shown that an optically based index derived from hyperspectral reflectance, called the Photochemical Reflectance Index (PRI), has considerable promise as a measure of vegetation radiation-use efficiency (?). This link to efficiency occurs because this index detects the response of xanthophyll cycle pigments to changing conditions (Gamon et al. 1990, 1997). Because these pigments are responsible for regulating absorbed light energy within the leaf, they provide a useful optical indicator of changing photosynthetic activity detectable with spectral reflectance. Thus, it is possible to express photosynthetic rates as a function of two terms, one that measures APAR (derived in part from NDVI), and one that measures efficiency (derived from PRI):

Photosynthetic rate = (PRI x APAR) (3a)

Photosynthetic rate = (PRI x (NDVI x PFD_PAR)) (3b)
 

A number of recent studies have shown that PRI can be significantly correlated with both photosynthetic radiation-use efficiency (?) and actual photosythetic rates at the leaf, canopy, and stand scales (Gamon et al. 1992, 1997, 2000, Penuelas et al. 1995, Stylinski et al. 2000). However, tests of this index at the landscape scale have rarely been attempted, in part because of the shortage of appropriate data sets. To validate this index, it is helpful to have an independent set of carbon flux (photosynthesis) measurements, which is often impossible at the landscape scale. Recently, eddy covariance has been developed as a means of sampling net carbon fluxes over large landscapes. By integrating the high-frequency covariance between carbon dioxide concentration and vertical velocity, this method provides a direct means of sampling instantaneous landscape fluxes. An additional challenge is that remote sensing typically samples in the spatial domain, whereas flux towers sample a single point in the temporal domain, making it difficult to compare the two data types (see figure 1). Typically expressed at half-hourly to daily intervals, but only for a single location, eddy covariance can provide flux measurements suitable for comparison with a small portion of an AVIRIS image. Ideally, to validate a map of CO₂ fluxes from AVIRIS, it would be best to have multiple flux towers sampling in a landscape, but this is normally not possible. The BOREAS data set, containing multiple AVIRIS scenes and data from multiple eddy covariance flux towers, provided a unique opportunity to explore the relationships between surface reflectance characteristics and CO₂ fluxes (figure 1).

Study site:

In 1993, the BOREAS project was undertaken in the boreal forest of central Canada, called the BOREAS modeling region (~500,000 km²). Two 50 x 50 km intensive study areas were selected the north and south ends of the modeling region. The northern study area (NSA) and the southern study area (SSA), as these two regions were named, were located near Thompson, Manitoba, and Prince Albert, Saskatchewan, respectively. They were about 500 km apart and each region contained flux towers located on relatively homogeneous vegetation and soil patches to measure water, vapor, heat and CO₂ fluxes using eddy covariance techniques. Our study utilized the CO₂ data collected during spring and fall seasons of 1994 from the SSA region flux towers. There were six flux towers in that region, covering sites dominated by black spruce (Picea mariana) , jack pine (Pinus banksiana)and aspen (Populus tremuloides) patches. Each flux site was relatively level and the forest stand was horizontally homogeneous throughout the area deemed as the flux footprint, a region extending over 1 km upwind.

Remotely sensed data and footprint area:

Airborne Visible InfraRed Imaging Spectrometer (AVIRIS) images from 1994 were used for this study. Images were available from April 19^th, July 21^st, and September 16^th of 1994, covering nine ‘flux-tower’ sites in total (Table 1). The fly-over time of the acquired images was approximately 11:30 a.m. local time. The original radiance images were corrected for atmospheric degradation using algorithms developed by Green et. al (1991). The atmospherically corrected images were then georeferenced, using a geo-corrected Landsat TM image of the area as a base map, and resampled to 30m pixel resolution.

The atmospheric- and geo-corrected AVIRIS images were used to produce NDVI and PRI images using the following equations:
 

(4)

and
 

(5)

In the above equations R represents reflectance, and the subscript B followed by a number represents the AVIRIS band represented by that number. AVIRIS bands #30, #47, #17 and #21 represent reflectance at wavelengths 660, 800, 531 and 570 nm respectively.

In order to make a scaled value of PRI (or sPRI) to represent it as a measure of efficiency, we did the following algebraic manipulation:
 

(6)

Footprint areas of the flux towers were determined based on the prevailing wind direction during the image acquisition time, as recorded by the tower instruments. Following the theoretical discussion of Kaharabata et al. (1997), a 300m X 900m rectangular area towards the upwind direction from each tower site was selected as an optimal footprint area. Average values of NDVI and PRI of those footprint areas were calculated for this study. Since NDVI itself can be expressed as a function of APAR, following the equation [3b], we then estimated a relative photosynthetic rate as a product of NDVI and scaled PRI:

Photosynthetic rate = (sPRI x NDVI) (7)

CO₂ flux data:

The ‘Tower-Flux’ science group of BOREAS research team collected CO₂ and other physiological fluxes from both SSA and NSA sites (http://eosims.esd.ornl.gov/BOREAS/bhs/Science_Groups.html). Although intended to operate continuously, most of the flux towers reported data sporadically during the 1994 season (Table 1). The eddy flux densities were determined by calculating the covariance between vertical velocity and scalar fluctuations (Baldocchi et al., 1988). Scalar fluctuations of CO₂ and other fluxes were computed in real-time, using a running mean removal method (McMillen, 1988). Then these flux data were averaged and stored at half-hour intervals. The measurement units were m mol m^-2 s^-1. Negative (-) flux values indicated net carbon uptake (i.e. photosynthesis) by the surface, and positive values indicated carbon loss to the atmosphere (i.e. respiration). We used CO₂ flux data from flux towers #2, #4, #5, #7, #9 and #11 for this study (Table 1), obtained through the courtesy of the respective "Tower-Flux" teams (http://eosims.esd.ornl.gov/BOREAS/boreas_home_page.html).

This half-hourly averaged CO₂ flux data showed considerable short-term fluctuation, which masked any trend over a diurnal period, even on a totally clear day. A further averaging of the half-hourly data to a two-hour running average produced a pattern smooth enough to show some diurnal trend (Figure 2). For comparison with AVIRIS imagery, we took an average value of CO₂ flux data from one hour before to one hour after the AVIRIS image acquisition time from the smoothed data as representative flux values for the tower region. This process was repeated for each of the nine tower sites and dates indicated in Table 1.

Averaged NDVI and PRI values from the footprint areas of nine tower flux sites were plotted against the two-hourly running average CO₂ flux values. The results are shown in Figure 3-A and -B respectively. The CO₂ flux data showed weak but significant correlations with both of these spectral indices, having an R² value of 0.57 with NDVI and that of 0.4 with PRI. NDVI and PRI were found to be not correlated with each other (Figure 3-C), indicating that in these landscapes NDVI and PRI detect two different aspects of green vegetation.
 
 

(8)

From equation 8, we derived CO₂ flux (midday net photosynthesis) maps from the AVIRIS images covering the tower sites. Representative flux maps of two tower site locations are shown in Figure 5. These two maps included the old black spruce and fen tower sites and their vicinity, and provided an empirically calibrated regional image of midday CO₂ fluxes.
 
 

Figure 4: Relation between CO₂ flux and a product of NDVI and scaled PRI from the nine tower sites in 1994. A linear relation was derived to relate CO₂ to the spectral indices, with an R² value of 0.78.

The primary conclusion of this study is that a simple, two-parameter model based on APAR (NDVI) and efficiency (PRI) can capture most of the variation in net photosynthetic fluxes across the boreal forest landscape. Although both parameters were significantly correlated with net CO₂ fluxes, NDVI had a more significant correlation than PRI, indicating that light absorption by green vegetation was the dominant factor driving CO₂ fluxes for these landscapes. The significant correlation between PRI and CO₂ fluxes are consistent with previous studies at the leaf and canopy scales indicating a relationship between this index and photosynthetic light use efficiency. The improved correlation with fluxes when both indices are used together (Figure 4) parallels other recent studies at the leaf and canopy scales that illustrate how a simple empirical model based on two optically derived parameters can readily capture much of the spatial variation in photosynthetic rates within sunflower stands (Gamon et al. 2000). To our knowledge, a direct correlation between this index and landscape fluxes has never before been possible, although previous studies of AVIRIS imagery have suggested the potential utility of this index as a measure of photosynthetic activity (Gamon et al. 1995a&b).

The approach presented here has many limitations that must be overcome if this method is to be used for routine mapping of CO₂ fluxes. The primary limitation derives from the empirical nature of this method, which requires the availability of flux tower data for calibration of AVIRIS imagery. In order to develop a process-based model of CO₂ flux assessment using hyperspectral imagery, there is a need of more coordinated flux and optical data collection from different biome types. Also the effects of view and solar zenith angles (bi-directional reflectance distribution) and atmospheric degradation on the remotely sensed data has to be quantified and modeled properly in order to use the remotely sensed data with confidence. Other studies not shown here suggest that errors in interpreting two-band indices can occur when pixels contain mixes of vegetation with other cover types, as was the case in our study. Despite these many limitations, the strong correlation between our simple two-parameter model and tower flux data indicates that considerable potential exists for mapping photosynthetic fluxes over large landscapes using narrow-band reflectance. Although not explored in this paper, it is likely that incorporation of improved surface cover maps (e.g. Fuentes et al., this volume) could lead to further refinements of this model.

Since AVIRIS images are not available for all seasons or all vegetation types, and atmospheric correction poses a problem due to lack of ground truthing during AVIRIS over flights, our next goal is to use tram- and light aircraft-based hyperspectral data collection platforms for the further development of this model. We are currently testing a hardware system and protocol to collect hyperspectral data from a chaparral stand near San Diego where an operational flux tower is also available to collect CO₂ and other flux data continuously.

Acknowledgements:

We thank the flux tower teams and BOREAS Information System (BORIS) staff for the data used in this study. This research was supported by a NASA BOREAS grant #NAG5-7401.

References: