1. Introduction
A strong interest is currently growing towards more extensive use of the red-edge and near infrared (NIR) spectral regions in vegetation remote sensing. New spaceborne sensors such as MultiSpectral Instrument (MSI) on board Sentinel-2 with special red-edge bands and Orbiting Carbon Observatory (OCO-2) for solar-induced chlorophyll fluorescence (SIF) retrievals based on in-filling of Fraunhofer lines in NIR spectral region have recently started to provide data with unprecedented quantity and quality [
1]. Substantial research has been conducted during last decade to enable SIF retrievals from reflectance at global scale [
2,
3], and the future ESA FLEX mission (planned for launch by 2023) will improve the spatial resolution of space borne SIF estimates to about 300 × 300 m
.
The majority of ecophysiological research uses Pulse Amplitude Modulated (PAM) chlorophyll a fluorimetry to assess plant physiological state via chlorophyll fluorescence. Among PAM chlorophyll a fluorescence parameters the steady-state fluorescence yield at ambient light intensity has the best resemblance to SIF. Previous research has shown that at single leaf level up to 90% of variability for steady-state fluorescence yield can be explained by leaf growth irradiance, nitrogen content and plant species [
4]. Absorbed light energy can be: (1) used for photosynthesis by photochemical energy conversion; (2) dissipated as heat; or (3) re-emitted as chlorophyll fluorescence. The first two processes are called photochemical and non-photochemical quenching, respectively, as both mechanisms reduce the chlorophyll fluorescence emission. If non-photochemical quenching (NPQ) is known or not changing, then chlorophyll fluorescence can be used to estimate photosynthesis. Active PAM chlorophyll a fluorimetry allows calculating parameters such as NPQ but not passive fluorescence measurement systems [
5]. Xanthophyll cycle is the protective mechanism that regulates NPQ to dissipate excess energy safely as heat and it also causes a subtle change at leaf level absorption spectra [
6]. Photochemical reflectance index (PRI) was originally constructed to assess the state of xanthophyll cycle via those spectral changes [
7]. However, further research has revealed that PRI is also affected by numerous other factors such as pools of carotenoids and chlorophylls [
8]. It has been shown that PRI at stand-level is not responding to the same physiological processes as at leaf level due to the effect of soil optical properties and canopy structure [
9]. Factors such as blue sky radiation caused by scattering in the atmosphere and forest structure via within-canopy illumination variations and non-physiological shadowing effects affect PRI at canopy level in boreal forests [
10,
11]. In addition, variations in direct and diffuse surface irradiance have strong influence on the relationship between SIF and PRI [
12]. To facilitate the future use of satellite derived SIF as an estimate of vegetation physiological state, leaf level studies on the relationships between PRI and different chlorophyll fluorescence parameters have been conducted [
13,
14].
Total chlorophyll fluorescence emission spectrum has two peaks at 685 nm (F685) and 740 nm (F740). It has been associated with the double-peak feature appearing in the red-edge spectral region of first derivative of vegetation reflectance and the ratio of the amplitude of the peaks by double-Gaussian fit has been used as an additional indicator of the photochemical activity of the plants [
15].
Many different vegetation indices have been developed since the beginning of spaceborne vegetation remote sensing in the 1960s and 1970s. The main principle of traditional vegetation indices is to compare the reflectance in the visible spectral region where strong absorption by pigments is prevailing with reflectance in near infrared (NIR) plateau where wavelength-independent scattering dominates. Different wavelengths from visible (400–700 nm) to far-red/red-edge spectral region (700–730 nm) have been used to track pigment absorption features [
16]. A review of different indices shows that in general vegetation indices that use only far-red spectral regions of longer wavelengths beyond 700 nm perform the best in predicting foliar chlorophyll content [
17]. Pigment absorption declines sharply after 700 nm where the absorption maximum of the reaction center of photosystem I (P700) is located and this spectral region of fast change is called “red-edge”. This sharp increase in reflectance (decline in absorption) is the most characteristic feature of vegetation spectra and special methods including different spectral fitting [
18] and linear extrapolation techniques for hyperspectral [
19] and multispectral [
20,
21] data have been developed to track and determine the precise location of the fastest change. The signal in Sentinel-2 MSI bands 5 and 6 of 15 nm spectral resolution compared to each other and to the neighbor bands allows estimating the position and steepness of the red edge. The spatial resolution of these two narrow spectral bands is less than other visible and NIR bands (bands 2, 3, 4 and 8 of 10 m pixels) in order to guarantee the required signal-to-noise ratio.
Many remote sensing studies deal with crops and sparse vegetation where the change in the amount of biomass simultaneously changes the relative contribution of reflectance signal originating from soil. Soil background optical properties have crucial influence on reflectance signal measured above the vegetation in such sparse canopies and therefore soil adjustment would be needed to extract the information about vegetation characteristics [
22]. The soil line concept has been used in formulation of many vegetation indices (e.g., Soil Adjusted Vegetation Index (SAVI) [
23], Modified Chlorophyll Absorption Ratio Index Improved (MCARI2) [
24], and Weighted (near-infrared-red) Difference Vegetation Index (WDVI) [
25]). However, this approach would not be appropriate in the case of multi-layered closed forest canopies in our current study, where background signal would originate from lower layers of trees, herbaceous understory or moss-layer.
Remote sensing measurements can be made at various scales ranging from single leaf level observations in laboratory to top-of-atmosphere (TOA) satellite observations [
26]. Airborne hyperspectral imagers CASI [
27] and APEX [
28] provide high quality spectral at-sensor radiance data. Such data or even non-calibrated digital numbers by new lightweight hyperspectral imagers [
29] can provide vegetation indices or perform classification of targets. For top-of-canopy spectral reflectance, simultaneous measurements of irradiance spectra are needed and, in the case of high-flying airborne measurements [
30], atmospheric correction is needed. Our study focused on the intermediate scale of so-called bottom-of-atmosphere (BOA)/top-of-canopy (TOC) measurements above mixed forest canopy. Low-level airborne spectral data are supported by simultaneous recording of irradiance spectra to convert airborne data to TOC spectral directional reflectance, and by forestry data provided by the National forestry database. The aim of the current study was to explore the inter-relationships between different spectral features of apparent top-of-canopy reflectance of hemiboreal mixed forest using measurements with medium or high spectral resolution, most commonly available for airborne spectral measurements and forest inventory variables.
2. Material and Methods
Airborne measurements of VisNIR directional reflectance were carried out over hemiboreal forests at the Järvselja test site in southeastern Estonia in summer 2010. The coordinates of the test site are 58.3°N, 27.3°E. The landscape at the test site is plain, and the ground height is 30–40 m asl. There are mixed forests which belong to the hemiboreal zone with moderately cool and moist climate and can be characterized as remote and rural with low anthropogenic disturbances. Stands are pure or mixed and composed mainly of silver birch (
Betula pendula), Scots pine (
Pinus sylvestris), Norway spruce (
Picea abies), common alder (
Alnus glutinosa), aspen (
Populus tremula), grey alder (
Alnus incana), and small-leaved lime (
Tilia cordata). A forest stand is a patch of homogeneous forest considering its species composition, age, tree height, tree density, and site type. A forest management inventory database including 1:10,000 map of stands is available for Järvselja Training and Experimental Forest district. The database is updated regularly. Several forest parameters such as species composition, age, breast-height diameter (cm), tree height H (m), basal area for the dominant and secondary layers G1 and G2 (m
/ha), stem volume for the dominant and secondary layer M1 and M2 (m
/ha), stem volume increment Z
(m
/ha/year), site type, etc. have been recorded for every stand. The minimal area of a stand is 0.1 ha according to the requirements by national forest inventory. Growth conditions at the study site range from poor where the site index H
(stand height at the stand age of 100 years) is less than 10 m to very good where H
can be over 35 m. About 75% of the site area is forests, natural grasslands, and pastures. A more detailed description of the test site was provided by Kuusk et al. [
31]. Clear-cut areas and stands younger than 15 years were removed from the current analyses. The final dataset reported here consists of 300 forest stands aged between 15 and 230 years. The stand areas are between 0.2 and 14.9 ha. The average stand age is 60 and median 54 years. Allometric estimate of tree layer leaf area index (LAI) ranged from 1 to 9 (mean and median 4.5).
Helicopter measurements of reflectance spectra in the spectral domain 350–1050 nm over the study area were carried out using UAVSpec3 spectrometer on 5 July 2010 (
Figure 1). The spectrometer UAVSpec3 is a fully autonomous lightweight spectrometer based on the 256-band near-infrared (NIR) enhanced version of the miniature spectrometer module Monolithic Miniature Spectrometer 1 (MMS-1) manufactured by Carl Zeiss Jena GmbH [
32]. The spectral resolution of the spectrometer is 10 nm and the spectral sampling interval is 3.3 nm. The spectrometer was mounted on the chassis of a Robinson R22 helicopter so that it was looking in the nadir direction during straight flight at constant speed. Average flight altitude was about 80–100 m above ground level and the flight speed was 60 km/h. The footprint of the field-of-view (FOV) of the spectrometer on the ground was about 2.5–3 m, and spectra were recorded at the frequency eight spectra per second. Measurements were carried out in direct sunlight at solar zenith angle (SZA) of approximately 40°. For the measurements of incident spectral radiation, the HR-1024 spectrometer by Spectra Vista Corporation equipped with a cosine receptor RCR/A124505 by Analytical Spectral Devices, Inc. was used. The SVC HR-1024 spectrometer has a spectral range of 350–2500 nm, 1024 spectral bands, and a bandwidth of about 1.5 nm in the wavelength range 350–1000 nm. Incoming spectral flux was measured at a nearby clearing.
Raw data from the airborne spectral sensor were corrected for dark signal temperature dependence and spectral stray light. The method suggested by Kuusk [
33] was used for dark signal correction. Stray light was corrected with the deconvolution method proposed by Kostkowski [
34] for spectral instruments. Instrument function of the spectral sensor was characterized using a double monochromator, as described by Kuusk et al. [
35]. The recorded digital numbers were converted to directional reflectance factor using simultaneous measurements of incoming spectral flux, calibration coefficients of every sensor element determined by measuring the calibrated grey reference panel SRT-20-120 by Labsphere Inc. [
36], and correcting recorded signals for dark current and stray light in the spectral sensor (
Figure 2),
Here,
is the spectral directional reflectance of the target at nadir (
°) at wavelength
measured at time moment
t;
and
are the signals of incoming flux during the target measurements and calibration, respectively;
and
are the signals of the UAVSpec sensor element which corresponds to the wavelength
; and
is the spectral reflectance factor of the reference panel. All the signals in Equation (
1) were corrected for dark current and the sensor signals
were corrected for stray light. The spectral resolution of the MMS-1 is less than that of the HR-1024; therefore, the recorded spectra of HR-1024 were filtered to the spectral resolution of UAVSpec3 using respective band-pass filter, and resampled to the wavelengths of UAVSpec3 before applying Equation (
1). As the distance between the top of forest stands and sensor was only 50–80 m, no atmospheric correction was applied. Measurements were carried out in stable illumination conditions. The calibration errors of spectrometers were reduced by using ratio of spectrometer signal in Equation (
1). The uncertainty of the reference
by Labsphere Inc. is less than 0.005 for the spectral range 300–2200 nm. In the following analysis included the stands over which at least 10 spectra were recorded. Spectra were sampled with the spatial step of 2.1 m on the flight transect. The footprint of the field-of-view of the UAVSpec at the flight height of 80 m is a circle of the diameter of 3 m. The range of the number of recorded spectra over a stand varied in general from 10 to 130, except for three mature stands over which several passes were done and the number of recorded spectra reaches 1300. Recorded spectra in the buffer zone of 8 m at the stand borders were not considered.
Average spectrum of the directional reflectance factor for every stand was calculated (
Figure 3A). The term Vegetation Index (VI) refers to a spectral transformation of two or more discrete spectral bands to enhance their sensitivity to vegetation properties. The most commonly used formulation of VI is the Normalized Difference Vegetation Index (NDVI).
It can be calculated for any spectral bands. The simplest form of VI is:
called Simple Ratio (SR).
Using spectral bands at 736 nm and 751 nm wavelengths suggested by Hallik et al. [
17], we found that SR (Equation (
4)) and NDVI (Equation (
5)) forms were almost linearly related to each other (
Figure 4C) and therefore we used only a single formulation of Equation (
4) in this study.
The other four vegetation indices used in this study (
[
7],
[
37],
[
38], and
[
20,
21]) were calculated as:
S2REP is the equation for the linear four-point interpolation approach based on the work of Guyot and Baret [
20] and Clevers et al. [
21] to calculate the red-edge inflection point using Sentinel-2 MSI spectral bands. The first derivative of reflectance spectra was calculated to find the true location of maximal change for each stand (
Figure 3B).
For the estimation of the sun-induced fluorescence, the spectral signatures of stands in the wavelength interval 752–765 nm were analyzed (
Figure 3C). At 761 nm, the incident sun radiation is strongly absorbed by oxygen absorption band O
-A in the atmosphere. In the recorded UAVSpec3 signal, the weak signal of reflected radiation is overlapped by the emitted radiation of sun-induced fluorescence [
39], therefore Equation (
1) returns the biased result
where
is the reflected radiance,
is the radiance of fluorescence, and
is the radiance of a non-absorbing Lambertian surface of no fluorescence at similar illumination conditions.
Contact laboratory measurements of reflectance spectra of landscape components—green leaves, tree stems and branches, and soil—using light source of no absorption lines show that all these spectra are smooth in the wavelength range 722–791 nm [
40]. Thus, in the case of no fluorescence, the apparent reflectance of forests were smooth too. This smooth spectrum is found by smoothing the recorded spectra in the wavelength range 722–791 nm by spline-approximation, omitting recorded values at 752–765 nm. This smoothed spectrum is called “continuum”. The difference of the recorded apparent reflectance and continuum near the absorption line at 761 nm is caused by the sun-induced fluorescence. The spectral resolution of UAVSpec3 is too low for recording the exact profile of the forest signal near 761 nm. The peak of apparent reflectance factor is spread by instrumental averaging over the spectral band 752–765 nm determined by the spectral sensitivity profile of the MMS-1. Therefore, the quantitative measure of the fluorescence contribution in the apparent reflectance is the area (not height) of the peak above the continuum near the absorption line at 761 nm,
where
is the wavelength interval 751.7–768.2 nm and
is the smoothed spectrum (continuum) (see
Figure 3C).
All statistical tests were performed with R version 3.4.4 (2018-03-15) [
41]. Spearman rank order correlation was used to assess the strength and direction of association between variables because some pairwise relationships were nonlinear. The significance level of 0.05 was used.
5. Conclusions
The red edge inflection point estimated in our study according to its mathematical definition as the location of the maximum of the first derivative of reflectance resulted in two discrete values around 717 nm and 727 nm when measured above mixed forest canopy. In clearcut areas, which were removed at the first stage of analysis as non-forest patches, a third possible location of maximum also appeared at 700 nm. No gradual shift was observed between peak locations estimated with this method. An alternative method of linear four-point interpolation for estimating red edge inflection point resulted in a continuous variable (called S2REP). The estimates of red edge inflection point with both methods were strongly related to each other but one method produced discrete and the other method continuous variables.
The in-filling of the O-A Fraunhofer line traditionally used for SIF estimation () was very strongly related to forest brightness in NIR spectral region ( = 0.91, p < 0.001) in our study. As a result, all variables which had correlations with single band reflectance factors at NIR plateau were also related to . NIR reflectance can be interpreted as a proxy for the proportion of sunlit foliage. The best predictor for forest age was single band reflectance at NIR region. The negative association between forest age and brightness in NIR region can be explained by the decrease in the proportion of sunlit foliage as forest grows. Biomass related forestry variables in our study (stem volume, basal area and tree height of dominant layer) were negatively correlated with single band reflectance factors at both visible and NIR spectral region, which is in accordance with the increased roughness of the forest canopy surface. The best predictor for allometric LAI was single band reflectance at red spectral region ( = 0.52, p < 0.001), outperforming all studied vegetation indices.
The new index R
/R
, which was found to be at leaf level the best predictor of chlorophyll content [
17], proved to work also at canopy level in the current study and was very strongly associated with red edge inflection point estimate S2REP (
= 0.99,
p < 0.001). The relationship between R
/R
and NDVI
was also strong (
= 0.95,
p < 0.001) suggesting that R
/R
could be useful alternative in those occasions where mathematical simplicity or narrow spectral range matters. No other vegetation index uses wavelengths so close to the O
-A Fraunhofer line.
We would recommend further exploration of the precise shape of the red-edge region with higher spectral resolution. The precise shape of the red-edge region from higher resolution spectral data could give much more information than simply discrete location of the red-edge inflection point and deserves to be studied much more. The current trend is towards the increase of the availability of very-high-resolution spectral data collected for SIF. Such information will soon be available even from the satellite level from ESA FLEX mission (scheduled launch in 2023). The number of peaks and the mechanism of peak jumps in red-edge derivative spectrum (particularly the interaction between canopy structure and physiological processes) should be elucidated considering the possible space-borne application in the future.