1. Sources of Variability in Canopy Spectra and the Convergent Properties of Plants
Funding From:
S.V. Ollinger, L. Lepine, H. Wicklein, F. Sullivan, M. Day
Earth Systems Research Center, University of New Hampshire, Durham, NH 03824, USA
Abstract
Remote sensing of canopy %N
Concentrations of nitrogen (N) in foliage are linked with a number of physiological and biogeochemical processes, making accurate detection central to our understanding of terrestrial carbon and nitrogen cycling over a range of scales. In this study, we explored the effects of spectral resolution, spatial resolution and radiometric resolution on estimates of forest canopy N concentration from remotely sensed data from airborne and satellite platforms. We conducted an exercise whereby PLS, simple and multiple regression calibrations to field-measured canopy N concentration for a series of forested sites were iteratively performed on (1) high spectral resolution Airborne Visible/InfraRed Imaging Spectrometer (AVIRIS) data sequentially degraded from 10nm to 90nm bandwidth; (2) Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) data simulated with AVIRIS data; and (3) actual data from Landsat and MODIS sensors. We observed virtually no loss of calibration accuracy with coarser bandwidths from AVIRIS, but instead saw accuracy decline as radiometric resolution and spatial resolution declined. We also found that regression models were driven primarily by reflectance in the near infrared (NIR) region, with little contribution from the visible or mid infrared regions. These results suggest that much of the variability in canopy N concentration is related to reflectance properties in the NIR region, and suggests a synergy among the scattering effects of leaf-, stem- and canopy-level traits that becomes accentuated in the near-infrared (NIR) region. This poses a serious challenge for remote detection of specific plant properties, but suggests an emergent property of ecosystems that results from optimization of plant form and function across multiple scales.
Site-specific forest canopy %N mapping
Canopy %N mapping across North American forests
Figure 2. (a) Location of study sites incorporated in previous studies that have examined CO2 uptake, canopy N and albedo at local, landscape and regional scales. Examples of our estimates of canopy nitrogen concentration are shown for several eastern U.S. sites within the AmeriFlux network, with estimates derived from PLS regression of field-measured canopy N against spectral reflectance data from NASA’s AVIRIS and Hyperion sensors (as described in Fig. 1). Data for forested sites here have been integrated to develop generalizeable methods for N detection (e.g. Martin et al. 2008). An example of a generalized PLS regression model is shown in (b), where measured %N is plotted against %N predicted with PLS regression for 155 forested plots across North America. The relationship is highly significant (p<0.0001) with a low standard error of calibration (0.18). Because this generalized PLS regression model was largely driven by NIR reflectance, we explored the possibility of estimating %N with NIR reflectance features from several sensors.
(e) PLS regression equation applied to AVIRIS image spectra to estimate landscape-scale foliar N concentration in Chibougamau, Quebec.
(b) Whole canopy %N determined for each plot
0.5
1.90 %N
plot
Howland, ME
Hubbard Brook, NH
Harvard Forest, MA
Duke Forest, NC
Austin Cary Memorial Forest, FL
Measured canopy %N
Canopy %N predicted with PLS regression
4-factor PLS regression of AVIRIS spectra with whole-canopy %N from155 plots across North America
(a)
(b)
(c)
(a) Foliage collection and plot characterization
Whole canopy %N related to AVIRIS spectra (above) with PLS regression (below).
(d)
Figure 1. Process for estimating whole canopy %N. Plot-level species composition and LAI are determined, and foliage is sampled for chemical analysis (a); whole canopy %N (% by foliar mass) (b) is related to AVIRIS spectra (c) with PLS regression; PLS regression equation is (d) applied to image spectra to estimate landscape-scale foliar N concentration (e).
Combined effects of multiple factors on whole-canopy reflectance ̶ Implications for nitrogen mapping
Spectral resolution
Vegetation indices and canopy %N
Only DVI was as closely related to %N as NIR reflectance alone
Here’s why
(a)
Influence of NIR region
(b)
One wide NIR band?
4.5
3.0
1.5 9 6 3
% Reflectance 24 16 8 15 10 5 60 40 20
1.0
2.0
Canopy %N
AVIRIS-simulated Landsat
Landsat
AVIRIS-simulated MODIS
MODIS
(a)
(b)
(c)
(d)
(e)
(f)
blue
green
red
NIR
mid-IR
Bandwidth R2
DVI
NDVI
NIR only
RVI
EVI
Figure 5. Observed relationships between canopy %N and reflectance from Landsat and MODIS bands. No significant relationships were observed between %N and visible reflectance (a–c) from simulated and actual Landsat bands and from simulated MODIS reflectance. Weak negative relationships were observed between %N and actual MODIS blue (r2 = 0.32, RMSE = 0.42), green (r2 = 0.13, RMSE =0.47) and red (r2 = 0.38, RMSE = 0.40) bands (a–c). Highly significant relationships (p<0.0001) were observed between %N and reflectance at the NIR band (d) for all data (simulated and actual), and weak correlations with the mid-IR bands (e-f). These observations illustrate why vegetation indices that represent some combination of NIR bands and visible bands do not tend to explain more variability in %N than NIR alone (see Fig. 4). 10 20 30 40 50
Reflectance
Wavelength (nm)
DVI = NIR - red
RVI = NIR / red
NIR - red
NIR + red
NDVI =
2.5 (NIR – red)
1 + NIR + 6 red – 7.5blue
EVI =
NIR reflectance
Measured canopy %N
Figure 4. The strength of relationships between vegetation indices and %N with progressively coarser bandwidth. Results from DVI were not statistically significantly different from NIR wavelength alone, with R2 greater than 0.80 for both across all band widths. Relationships between %N and RVI, NDVI and EVI were all weak, but strengthened with coarser bandwidth. This is likely a result of expanding the width of the red region so it includes the red-edge.
Simple regression of canopy %N and AVIRIS reflectance averaged over the wavelengths from nm
Increasing NIR reflectance with increasing canopy %N
Figure 3. PLS factor loadings (e.g. from relationship shown in Fig. 2b) were most influenced by NIR bands. This likely reflects the pattern of increasing NIR reflectance with increasing %N observed in (a) above, where average AVIRIS spectra were plotted for our study sites grouped into 9 classes of %N. The influence of the NIR region is further demonstrated in (b), where AVIRIS reflectance was averaged over the wavelengths ranging from ~ nm and entered into a simple linear regression on %N for 155 plots. The relationship between %N and this one 90nm-wide band was highly significant (p<0.0001), with a reasonably low standard error (RMSE = 0.24).
AVIRIS-simulated Landsat
Landsat
AVIRIS-simulated MODIS
MODIS
Spatial resolution: Comparison of AVIRIS and forthcoming HyspIRI
Factors affecting leaf and canopy reflectance
Functional convergence among spectrally important plant traits
Variability in %N is still captured in a 60m pixel
(b)
(a)
60m pixel
R2 = 0.76
SE = 0.24
18m pixel
R2 = 0.84
SE = 0.21
Potential importance of interrelated plant traits that affect radiation scattering over scales ranging from cells to canopies.
Comparison of PLS calibration of AVIRIS imagery degraded from 17m to 60m pixel
Figure 8. Examples of known associations that illustrate convergence of spectrally important plant traits are shown (a); idealized relationships among other variables that exhibit some degree of convergence and that are known to be related to NIR reflectance are (b) (Ollinger 2011).
Figure 7. Leaf (a) and canopy (b) reflectance spectra predicted by the PROSPECT and SAIL models, generated using a range of values for Chl concentration, dry matter content, EWT, and the structure parameter N; and LAI and leaf angle distribution (LAD) for whole-canopy reflectance (Ollinger 2011).
The factors that most affect reflectance are those that are most difficult to measure or estimate … And these and other important factors have potentially strong interrelations.
Figure 6. Demonstration of the effect of spatial resolution on estimates of canopy %N. Here, a comparison of PLS regression results with 17m vs. 60m pixels. We lose very little predictive power in N detection (a) and we still capture the variability in patterns of canopy %N across the landscape (b).