Links between light availability and spectral properties of forest floor in European forests

Remote sensing using spectral data has been commonly applied to retrieve tree layer properties while the monitoring of forest floor remains a less studied topic. We investigated the links between light availability at forest floor, and forest floor ’ s spectral reflectance properties (350 – 2500 nm) and fractional cover across boreal and temperate Europe. We hypothesized that tree canopy structure (and thus, light availability at forest floor) is linked not only to the vegetation composition of forest floor, as has been shown previously, but also to forest floor ’ s spectral reflectance properties, and that these relationships differ between forest biomes. Data were collected in situ from a total of 67 forest stands in southern boreal, hemiboreal, temperate floodplain, and temperate mountain sites. The variation of light availability at forest floor was linked to both the forest floor ’ s composition and spectral reflectance properties. Each study site exhibited site-specific spectral features and a different mean reflectance spectrum. Openness in tree canopies was related to an increase in the fractional cover of vascular plants and to a decrease of plant litter, consequently enhancing the forest floors ’ spectral absorptance features in the red and shortwave-infrared wavelengths, as well as reflectance in the near-infrared region. Also, the variations of normalized difference index values and red edge positions as functions light availability at forest floor and forest floor ’ s composition were different for each site. Our results suggest that incorporating biome-specific relationships between tree canopy structure and forest floor reflectance properties would improve interpretation of optical remote sensing data. The measurement data are openly available.


Introduction
Assessment of forest ecosystem services requires up-to-date data on forest attributes that are not limited to the woody components of a forest.For several decades, remote sensing techniques have been used to retrieve information on the overstory layer (tree canopy) of a forest, and more recently, also to assess ecosystem services (e.g., Andrew et al. 2014;Vauhkonen 2018).However, remote sensing of forest floor has been an understudied topic even though forest floors' vegetation and soil play key roles in forest productivity (e.g., Kolari et al. 2006), promote forests' biodiversity, and also have economic value (e.g., berry production by shrubs and dwarf shrubs; e.g., Kangas 1999).Previous research has shown that the composition of forest understory is strongly correlated with structural forest canopy measures in European boreal (Majasalmi and Rautiainen 2020) and North American temperate (Cole et al., 2017) forests, and that the spectral characteristics of a forest floor are linked to its fractional cover and forest site type in European boreal (Rautiainen et al., 2011;Forsström et al., 2021) and hemiboreal (Nikopensius et al., 2015) forests, as well as to understory leaf area index (George et al., 2021).However, a comprehensive investigation of the relationships between forest canopy structure and forest floor spectral reflectance properties across biomes has not been previously conducted.Such studies are needed to examine how universal the relationships are.The tree layer has an impact on understory vegetation: it influences the light conditions (Tonteri et al., 2016;Hovi and Rautiainen, 2020;Hase et al., 2022) and litterfall and nutrient cycles (e.g., Nilsson and Wardle 2005;Ukonmaanaho et al. 2008).A more complete view of how hyperspectral remote sensing could be used to assess forest ecosystems can only be obtained if the relationship between tree-and forest floor layers is understood in detail in the context of remote sensing.
The contribution of forest floor to the remotely sensed forest reflectance has been recognized to be notable (Spanner et al., 1990;Chen and Cihlar 1996;Eriksson et al., 2006;Rautiainen and Lukeš 2015;Hase et al., 2022) and attempts to estimate properties of forest floor from remote sensing data have been made (e.g., Markiet and Mõttus 2020;Pisek et al. 2012;Pisek et al. 2015Pisek et al. ,2016Pisek et al. , 2021)).Previous studies on forest floor spectra have shown that in European boreal (Rautiainen et al., 2011;Forsström et al., 2019), hemiboreal (Nikopensius et al., 2015), and temperate (Hase et al., 2022) forests, forest floors exhibit seasonally dynamic spectral signatures that depend on the vegetation cover and fertility of the site.However, preceding studies have been restricted to a small number of stands or samples (Rautiainen et al., 2011;Nikopensius et al., 2015;Hase et al., 2022), are made in laboratory under artificial illumination conditions (Peltoniemi et al., 2005;Forsström et al., 2019;Kuusinen et al., 2020), or have used instrumentation with limited spectral range within 350-1000 nm (VNIR) (Miller et al., 1997;Rautiainen et al., 2011;Pisek et al., 2021).Extension of the spectral range of in situ measurements beyond VNIR into shortwave-infrared (SWIR) wavelengths (1000-2500 nm) could provide valuable insights on forest moisture condition, and further soil fertility and forest production (Sikström and Hökkä 2016), and also be applicable to interpretation of data from new and forthcoming satellite missions (e.g., Sentinel-2, Landsat 9, PRISMA and EnMAP) which measure also in SWIR wavelengths.
In this paper, our goal is to bridge the gap between remote sensing and light regime studies related to forest floors, and to create a fundamental scientific understanding of the relationships between tree canopy structure, and spectral properties and composition of forest floor.We hypothesize that tree canopy structure, and hence light availability at forest floor, is connected not only to the composition of forest floor vegetation (as has been reported in previous studies) but also its spectral reflectance properties.We report results based on field measurements conducted in study sites representing different forest ecosystems across boreal and temperate Europe.Specifically, we examine how light availability at forest floor level is linked to fractional cover estimates, reflectance spectra and spectral indices of forest floor, be it covered by vegetation, lichens, or litter.The data reported in this paper are available as an open data set (Forsström et al., 2023a, Forsström et al., 2023b).In the future, our results and data can serve as the basis for developing methods for remotely sensed estimation of forest floor properties and help provide a more complete view of forest ecosystems.

Study sites
We made field measurements in different types of European boreal and temperate forests in Finland (2018Finland ( , 2019)), Estonia (2020), and the Czech Republic (2019) near peak growing season (Fig. 1).The study sites were in structurally different forests, and the study stands were selected with the aim of covering maximal tree canopy structure (plant area index, tree species, canopy cover) and understory variation within the entire dataset (all sites included).Thus, the study stands do not represent a random (unbiased) sample of the forests in each study area.In particular, the age and height distributions are not balanced: young forests (tree height < 10 m) are overrepresented in the temperate mountain (n = 3) and hemiboreal sites (n = 3), compared to the southern boreal (n = 1) and temperate floodplain sites (n = 0).Altogether, we collected data from 67 study stands (Table 1).
Our northernmost study site was located in Finland, near the Hyytiälä Forestry Field Station (61 • 51 ′ N, 24 • 18 ′ E) (Fig. 1).This southern boreal study site comprised of 36 upland (mineral soil) study stands, dominated by coniferous Norway spruce (Picea abies (L.) Karst) and Scots pine (Pinus sylvestris L.), and to a lesser degree broadleaf birches (Betula pendula Roth, Betula pubescens Ehrh.).The forest floor composed graminoids, mosses, dwarf shrubs, lichens, and plant litter (meaning dead needles, leaves, bark, cones, twigs) (Table 1).Variation in the tree layer and forest floor compositions between study stands was considered large in the southern boreal site and has been previously linked to stand level fertility characteristics (see Majasalmi and Rautiainen 2020).
The hemiboreal study site in Estonia was located close to the Järvselja Training and Experimental Forestry District (58 • 17 ′ N, 27 • 19 ′ E) in the hemiboreal transition zone between boreal and temperate regions (Fig. 1).Our 13 study stands were in forested land on mineral or peatland soils, dominated by broadleaf birches (Betula pendula, Betula pubescens), black alders (Alnus glutinosa (L.) Gaertn.), and European aspen (Populus tremula L.), with coniferous Norway spruces and Scots pines being less common.Most of the forest floor was composed of tall herbaceous plants, but also dwarf shrubs and moss were present in less fertile stands.Ground that was not covered by green vegetation had mostly intact plant litter (Table 1).
Two of our study sites represented different types of temperate forests.In Bílý Kříž, the Czech Republic (49 • 30 ′ N, 18 • 32 ′ E) our study site was in high elevation mountain forest (700-950 m a.s.l.) and included eight study stands; we refer this site as the temperate mountain site (Fig. 1).The high-reaching tree canopies in our study stands in the temperate mountain site were dominated by Norway spruces, with small amounts of European beech (Fagus sylvatica L.) and other broadleaf species.The forest floor was composed of low-growing dwarf shrubs, graminoids, and moss, and needle litter, mostly covered by dense tree canopies (Table 1).The study site in Lanžhot, the Czech Republic (48 • 41 ′ N, 16 • 57 ′ E), on the other hand, was in a lowland floodplain forest and included 10 study stands (Fig. 1).We refer to this site as the temperate floodplain site.The stands in the floodplain site had closed tree canopies, composed of a large diversity of broadleaf species, mainly oaks (Quercus robur L., Quercus petraea (Matt.)Liebl.), ashes (Fraxinus excelsior L., F. angustifolia Vahl.),European hornbeam (Carpinus betulus L.), hedge maple (Acer campestre L.), white poplar (Populus alba L.), European aspen (Populus tremula L.), and littleleaf linden (Tilia cordata Mill.).A major fraction of the forest floor was covered in leaf litter in various stages of decomposition and low-growing shrubs, while herbaceous plants were observed in areas where the tree canopy was more open (Table 1).
In all 67 study stands, we measured forest floor reflectance spectra (Section 2.2), tree canopy diffuse non-interceptance (DIFN) (characterizes light availability at the forest floor level; Section 2.3), and fractional cover of the forest floor (Section 2.4).We conducted also basic forest inventory to characterize the stands by tree height and stand density, and to classify each stand according to so-called dominant tree type as either pine, spruce, broadleaf, or mixed stand, based on the tree species composition and an 80% threshold for dominance.For three temperate mountain stands, firs (Abies alba; max 16% per stand) were counted as spruce trees during tree type classification (Table 1).

Forest floor spectra
Forest floor reflectance spectra were measured along a preestablished, 11 m long East-West oriented transect under diffuse illumination conditions (overcast sky, or close to dusk or dawn) in each stand.We used an ASD FieldSpec4 standard-res spectrometer (serial number 18,456) in a backpack set-up with a pistol grip and bare fiber fore optics to record the relative spectral radiation over 350-2500 nm wavelength range.The instrument has spectral resolution of 3 nm at wavelengths up to 1000 and 10 nm at wavelengths above 1000 nm, and it interpolates and outputs the spectra at 1 nm interval.For a measurement position, the spectral signal of the forest floor was recorded from a circular spot (approximately 40-50 cm in diameter) on the ground by holding the pistol grip at arm's length, with the sensor optics (25 • opening angle) leveled and pointing downwards.The exact sensor footprint depended on prevailing microtopography of the ground surface and understory height.We made frequent white reference measurements at regular intervals along the length of a transect from a reference panel (Spectralon®, 99% nominal reflectance) and the reference measurements were made next to the forest floor measurement positions while holding the panel horizontal at knee height.
Our sampling procedure consisted of 15 forest floor spectral measurements from evenly spaced sampling positions (separated by ~ 80 cm) along the transect, as well as reference measurements at every third P.R. Forsström et al. and at the last sampling position.Further, we recorded the spectrometer's dark current.To guarantee an adequate signal level during the measurement of a transect in each stand, the spectrometer was optimized before measurement of each transect.However, for some individual transects, the integration time was adjusted manually downwards (to a predefined maximum of 2.18 s) to ensure a reasonable overall measurement time per transect.This was a compromise between adequate signal level and rapidly changing light conditions in measurements made in the morning or evening.The number of observations was 15 for one averaged spectrum from each measurement position.The spectrometer was allowed to warm-up outdoors at least 30 min prior to the measurements.
The forest floor spectral data were recorded in sensor-specific intensity values, referred to as digital numbers (DN).These were converted to hemispherical-conical reflectance factors (HCRF, hereafter referred to simply as reflectance), according to the illumination and viewing geometries of the measurements (Schaepman-Strub et al., 2006).The forest floor reflectance for each measurement position was calculated as where DNs are respectively the wavelength dependent (λ) signal values from forest floor (FF), white reference panel (WR) measurements at the specific times corresponding to the forest floor's spectral measurement (t FF ), and the average of dark current (DC) measurements for the transect.R ref is the reflectance of the panel used as a correction term to account for the non-ideality of the panel's reflectance properties.Linear interpolation in time was used to derive white reference at the times of measurements of the forest floor spectra.Data in the spectral regions, where the measured spectra were noisy due to atmospheric attenuation of incoming solar radiation (1336-1504, 1781-2094, 2316-2500 nm), were excluded from the analysis.We also removed spectral data, either completely (350-2500 nm, due to measurement errors or missing data) or in the SWIR region (2095-2500 nm, due to high noise) in a small number measurement positions.For two southern boreal stands, all data in SWIR (2095-2500 nm) were removed due to high noise.
A mean spectral reflectance for each study stand was calculated from the 15 reflectance measurements, and for each biome from the means of the study stands.We smoothed the mean spectral reflectance and deviation data for each study stand and study site separately to reduce the level of random noise using Savitzky-Golay filtering algorithm (2nd order polynomial, window size of 41 nm) in Matlab.Site-specific spectral standard deviation and coefficient of variation were calculated from the unfiltered data.
Further, we calculated narrowband normalized difference indices (NDI) as and the coefficient of determination (R 2 ) between NDI, and light availability (DIFN) and fractional covers of three classes (vascular plants, non-vascular plants, and intact plant litter) for all wavelength combinations of the hyperspectral data.This was done to assess the dependency of NDI on DIFN and forest floor's composition.Finally, we examined the red edge features of the forest floor.We located the red edge position (REP, i.e., wavelength position of the maximum amplitude of the first derivative reflectance curve between 680 and 740 nm) and calculated two red edge shape features, namely, peak amplitude (dr re , i.e., amplitude of the first derivative reflectance curve at REP) and area (Σdr 680-740 nm , i.e., sum of the amplitudes of the first derivative reflectance curve between 680 and 740 nm) (Filella and Peñuelas, 1994) for each stand.We then assessed the relationships between these features, and DIFN and fractional cover of vascular and non-vascular plants, and intact plant litter.The changes in REP and the shape features of the red edge for green vegetation have been attributed to changes in the leaf and foliar chemistry, such as chlorophyll and nitrogen (Horler et al., 1983;Curran et al., 1990;Vogelmann et al., 1993;Filella and Peñuelas, 1994;Gitelson et al., 1996;Mutanga and Skidmore, 2007), water content and water stress (Horler et al., 1983;Maimaitiyiming et al., 2017), canopy cover (Curran et al., 1990;Clevers et al., 2001;Pu et al., 2003) and biomass (Horler et al., 1983;Vogelmann et al., 1993), as well as forest floor's seasonality (Rautiainen et al., 2011;Nikopensius et al., 2015;Forsström et al., 2019).

Forest floor light environment
Hemispherical photos were used to assess the light environment at the forest floor for each study stand.The diffuse non-interceptance of the tree canopy (DIFN, a proxy for the forest floor's light environment) and the effective plant area index of the tree canopy (PAI e ) were calculated from the photos.A total of five hemispherical photos were acquired from each study stand at the first, middle, and last spectral measurement positions, and at two positions 5.5 m in south and north directions from the center of the transect.
Our camera system included a Nikon D5000 camera body and a Sigma 1:2.8 DC HSM Circular Fisheye 4.5 mm lens.Before measurements, the lens aperture was set between F5 and F11, and the sensor sensitivity to a low ISO number to achieve best possible image quality.The photos were taken during diffuse sky conditions (based on visual judgement) to ensure homogenous image quality.To normalize camera exposure under different types of light environments (due to varying canopy cover, thickness of clouds, and time of day), we used stand-and illumination-specific exposure times to achieve maximum range of 8-bit values.To ensure that a good-quality image was always available, we used the autobracketing function of the camera to take also photographs with exposure times doubled and halved from the original setting.During the measurements, the camera was fixed to a tripod (at height of 1.3-1.5 m, or at 1 m in stands where the mean stem diameter at breast height was less than 10 cm) and the surface of the lens (hood) was leveled horizontal.The camera was aligned so that the top of the image frame faced in compass North direction.Furthermore, we used the selftimer function or remote trigger to avoid camera shake.
Hemispherical photographs were used to calculate DIFN and PAI e for each study stand (Table 1), based on the ratios between tree canopy and sky pixels in concentric circular areas.Our processing chain for the photos incorporated an automatic threshold method based on edgedetection, proposed by Nobis and Hunziker (2005) and implemented later in a code by Korhonen and Heikkinen (2009).The algorithm evaluates intensity contrast between neighboring gray pixels (8-bit) and aims to find an optimal threshold value (0-255) for global binarization of the image (sky pixels get value 1 and tree canopy pixels value 0).For processing, we extracted the blue channel of the RGB-photos for maximum contrast, as blue light is pronouncedly affected by the atmospheric Rayleigh scattering (sky appears bright) and strongly absorbed by green vegetation (tree canopy appears dark).Histograms of the three differently exposed gray scale images obtained from bracketing (-1, 0, +1 EV) from each measurement position were examined and the image exhibiting the most distinct bimodal frequency distribution, meaning that the histogram showed well-separated peaks within the data range without being saturated was selected for further calculating the tree canopy gap fractions.We segmented all the selected images into five concentric sampling rings with following zenithal annuli (or view zenith angle range): 0-15  divided by the number of all pixels, and DIFN and PAI e were calculated from the gap fractions following the methodology presented in LAI-2200 Plant Canopy Analyzer instructions manual (LI-COR 2012).

Forest floor fractional cover
For each study stand, the forest floor's fractional cover was visually estimated from four nadir-view RGB photographs, taken at every fourth spectral measurement spot along the transect.Each photo covered a 1 × 1 m vegetation quadrat frame on the ground.The photos were taken handheld at a height of approximately 150 cm above the forest floor using sufficiently high shutter speeds to avoid camera shake.
Classification of the forest floor fractional covers was done manually for the following classes: vascular plants, non-vascular plants (i.e., moss), lichen, intact plant litter (leaves, needles, sticks, cones, etc.), and decomposed plant litter.First, for each photo, the area framed by the quadrat was rectified and cropped to obtain a square-shaped image of the quadrat.Next, each quadrat image was overlaid with a 10 × 10 grid, so that a single sub-section of the grid represented 1% of the total image area.Then, each sub-section was classified as one of the fractional cover classes according to the dominant material and the fractional cover of each class in the image was determined as the sum of the sub-section classes.Finally, the fractional cover of each class for a study stand was calculated as the mean of the four images.A single person was responsible for the visual classification to ensure consistent results.

Results and discussion
Our four study sites represented different types of forest biomes with distinct tree layer and forest floor structures and composition (Table 1).In general, variations in the light availability at the forest floor (DIFN), its fractional cover classes, and reflectance spectra were interconnected.The results are organized as follows: in Section 3.1, we show how the tree canopy DIFN and forest floor's cover fractions varied between the four study sites and discuss their interlinks; in Section 3.2, we show the measured forest floor spectra and their variation both within and between the four sites, and discuss the impact of forest floor composition on its reflectance properties; and in Section 3.3, we examine the links between tree canopy DIFN and forest floor spectral properties.

Light availability and composition of forest floor in the study sites
The light availability at the forest floor was evaluated using DIFN, which ranged from 0.01 (closed canopy) to 0.83 (mostly open canopy) for all study sites (Fig. 2A).The temperate floodplain site had the most closed tree canopies (mean DIFN = 0.09), while the hemiboreal site had, on average, the most open canopies (mean DIFN = 0.34).In the southern boreal site, the tree canopies were substantially more open (mean DIFN = 0.26) than those in the temperate floodplain site but less so compared to the temperate mountain site's canopies (mean DIFN = 0.29).The spruce-dominated temperate mountain stands exhibited the largest within-site variation of DIFN and had two stands with particularly open tree canopies (DIFN = 0.83 and DIFN = 0.66).At the other end, the relatively dense broadleaf tree canopies that dominated the temperate floodplain site exhibited the smallest within-site variation of DIFN.
As background information, we also present tree canopy PAI e for each stand (Fig. 2B).PAI e accounts for both the silhouettes of foliage and woody elements, and was calculated from the same gap fraction measurements as DIFN.Based on Beer-Lambert law, there exists an exponentially decreasing relationship between PAI e and DIFN where small DIFN is related to very large PAI e .
On average, the hemiboreal study site had the largest fraction of vascular plants (0.63) and the smallest fraction of intact plant litter (0.27) on the forest floor, while the temperate floodplain site had the smallest fractional cover of vascular plants (0.28) and largest fractional cover of both intact and decomposed plant litter (0.55, 0.16, respectively) (Fig. 2C).The southern boreal site, on the other hand, had large fractions of both vascular (0.45) and non-vascular plants (0.21), and was also the only study site with lichen cover (0.02) (Fig. 2C).Lichen was notably present only in pine stands.Finally, in the spruce dominated temperate mountain site, nearly half of the forest floor was covered by either vascular (0.39) and non-vascular plants (0.05), and the rest was intact plant litter (0.56) (Fig. 2C).
Next, we analysed the relationships between light availability at the forest floor level (DIFN) and three forest floor fractional cover classes common to all study sites (Fig. 3).In general, an increase in DIFN promoted an increase in the vascular plant cover and a decrease in the plant litter cover, while the effect of DIFN on non-vascular plant cover remained unclear.The strength and shape of these relationships varied between the study sites.A linear relationship between DIFN (x) and vascular plant cover (y) was evident in the temperate floodplain (R 2 = 0.79 for y = 2.98x, n = 8) and temperate mountain site data (R 2 = 0.75 for y = 1.08x + 0.09, n = 7), while a non-linear model fitted better to the hemiboreal site data (e.g., R 2 = 0.44 for y = -1.05x−0.08 + 1.85, n = 13).
No model provided a good fit to the southern boreal site data (R 2 = 0.01 for best linear fit, n = 36).Similarly, the decrease of intact plant litter coverage (y) as a function of DIFN (x) was fairly linear in the temperate mountain (R 2 = 0.66 for y = -1.09x+ 0.87, n = 8) and floodplain (R 2 = 0.53 for y = -2.14x+ 0.75, n = 10) sites, while less so in the hemiboreal site (R 2 = 0.28 for y = -0.34x+ 0.39, n = 13) and neglectably so in the southern boreal site (R 2 = 0.03 for best linear fit, n = 36).For the nonvascular plant cover type, the relationships with DIFN were considered very weak for the southern boreal, temperate mountain, and hemiboreal sites (R 2 = 0.01, n = 33; R 2 = 0.04, n = 5; R 2 = 0.01, n = 10 for best linear fits, respectively), and unclear for the temperate floodplain site which had only two stands with fractional cover of non-vascular plants ≥ 0.01.Our results on the forest floor fractional covers of vascular and non-vascular plants, and intact plant litter in each study site (Fig. 2A) and their relationships with tree canopy structure (Fig. 3) were similar to those reported previously by Majasalmi and Rautiainen (2020) from the same southern boreal forest area.However, for the southern boreal site specifically, the preceding study reported strong or moderate correlations (r) between tree canopy openness, and forest floor's vascular (referred to as 'upper story' in Majasalmi and Rautiainen ( 2020)) (r = 0.28) as well as litter covers (r = -0.19),which were practically non-existent in our study.This may have been because our sampling in the southern boreal site included a relatively smaller number of open-canopy stands (DIFN > 0.5) compared to the preceding study.
When we examined the data collectively from all sites for the relationships between DIFN and vascular plant cover by dominant tree types (i.e., pine, spruce, broadleaf, and mixed), we observed moderate linear relationship in broadleaf stands (R 2 = 0.42 for y = 1.02x + 0.30, n = 20), and weaker linear relationships in mixed tree type (R 2 =0.37 for y = 0.69x + 0.32, n = 9) and spruce (R 2 =0.28 for y = 0.78x + 0.25) dominated stands.Similarly for intact litter cover, the relationship was linear and weak for broadleaf (R 2 =0.31 for y = -0.78x+ 0.59, n = 22) and mixed (R 2 = 0.30 for y = -0.49x+ 0.48, n = 9) stands and very weak for pine stands (R 2 =0.14, n = 17).For spruce stands the best fits were non-linear (e.g., R 2 = 0.43 for 0.09x −0.71 , n = 19).For the non-vascular cover, none of the tree types had a clear relationship, either because of scattered or unbalanced data.

Spectral properties of forest floor
To characterize the spectral variation of the forest floors in our study sites, and to assess if future hyperspectral sensors could have potential in monitoring forest floor vegetation, we first examined the forest floors' reflectances using continuous representations of the data in full spectral resolution and range, i.e., as hyperspectral data (Fig. 4A-C).The shapes of the spectral signatures varied between the biomes (or sites) and the observed differences of the spectral features were linked to differences in the forest floor compositions (Fig. 2C).Both the southern boreal and hemiboreal sites had large fractional covers of vascular or non-vascular plants and exhibited notably strong reflectance peaks in the green wavelengths (~520 -580 nm).However, of all sites, only the hemiboreal site showed a strong absorption peak in the red wavelengths (~675 nm), indicating a high chlorophyll content (e.g., Gitelson et al., 2003).The relatively lower absorption (high reflectance) features of the southern boreal site in the red wavelengths were probably due to spectral influence of moss (Vogelmann and Moss, 1993;Bubier et al., 1997;Peltoniemi et al., 2005) and lichen (Peltoniemi et al., 2005;Kuusinen et al., 2020), both of which are known to be bright throughout the visible part of the spectrum.The hemiboreal site exhibited also distinctly sharp increase of reflectance along the red-edge (690 -750 nm) up to NIR-shoulder (~750 nm), and highest reflectance in NIR, with two spectral troughs induced by the water content (at ~965 nm and ~1190 nm) (Knipling, 1970;Curran et al. 1989).The spectral troughs were present also for the southern boreal, hemiboreal, and temperate mountain site spectra, while less so for the temperate floodplain site.In SWIR, all four study sites exhibited similarly bell-shaped spectral curves, confined by the (atmospheric) water absorption bands.The hemiboreal site had the darkest forest floors in most of the wavelengths in SWIR, probably due to largest fractional cover of vascular plants and thus, largest water content, and temperate floodplain was the brightest, indicating a relatively dry forest floor (e.g., Hunt and Rock 1989;Carter 1991;Gao 1996;Nagler et al. 2000;Seelig et al. 2008;Ustin et al., 2012).The forest floors of the two temperate sites were similar with each other in the visible wavelengths and partly in NIR but different at the longer wavelengths of the NIR region (> 1000 nm) and in SWIR: the forest floor of the temperate floodplain site was brighter in SWIR, indicating a lower forest floor moisture content.
In most wavelengths, the within-site variation of the forest floor spectra was largest in the temperate mountain site and smallest in the temperate floodplain site (Fig. 4B, C).However, exceptions to this were noted in parts of the red edge and NIR region, where the variation was largest in the hemiboreal site and smallest in the southern boreal site.The temperate mountain site exhibited a distinctly large within-site variation at and near the spectral absorption peak of chlorophyll (~630-680 nm), as well as in SWIR, probably due to large variation of fractional covers of vascular and intact plant litter between the stands (Table 1).In the hemiboreal site, the large variation of NIR spectra may be linked to notably tall understory vegetation and its deep shadows in some of the stands.Interestingly, neither the variation in the forest floor composition nor in the dominant tree type between stands (Fig. 3) induced a notable variation in the forest floor spectra.This was the most evident for the southern boreal site where the standard deviation of forest floor reflectance spectra was consistently low compared to other sites, despite exhibiting most diverse forest floor composition and tree canopy characteristics (Table 1).
We also examined the spectral data in multispectral form, resampled to selected spectral bands of the Sentinel-2 MSI sensor (Fig. 4D).Following from the hyperspectral data, forest floors in all sites exhibited low spectral reflectance (high absorptance) in the visible ultra-blue, blue, green, and red bands, an increase in reflectance in the red edge bands to NIR bands, and a decrease of reflectance in the SWIR bands.The hemiboreal site with the largest fractional cover of vascular plants of the four sites showed an enhanced absorption feature (low reflectance) in both the red and SWIR bands, as well as high reflectance features in the NIR bands.Such spectral features have been associated with those of green-leaved vegetation (Gates et al. 1965;and later, e.g., Hovi et al. 2017) and consequently, forest floors during peak growing season with green understory in boreal (Rautiainen et al., 2011) and hemiboreal forests (Kuusk et al., 2004;Nikopensius et al., 2015).In contrast, for the temperate floodplain and mountain sites, we observed gradually increasing reflectances throughout the spectrum from ultra-blue and blue to NIR bands, and weak absorptance features in the red band, corresponding to known spectral characteristics of plant litter and soil (Nagler et al., 2000;Conforti et al., 2018;Seelig et al., 2008;Forsström et al., 2019;Tian et al., 2021), as well as many lichen species (Kuusinen et al., 2020).These results indicate that coarse discrimination of forest floor composition is possible using multispectral information.
Our results for boreal and hemiboreal sites' forest floor mean spectra were supported by earlier findings by Kuusk et al. (2004), Rautiainen et al. (2011), andNikopensius et al. (2015).For the temperate sites our results were different from those reported by Hase et al. (2022) near peak growing season: namely, while both studies showed similar monotonic trends of increasing reflectance from visible to NIR wavelengths, our results showed lower NIR reflectances for both temperate sites.This was probably due to less green forest floor compositions in our study sites (i.e., larger fractional covers of intact and decomposed litter).
Analysis of vegetation indices for individual study sites showed that there were a wide range of NDIs available for predicting vascular cover across the measured spectral range for hemiboreal, and temperate mountain and floodplain sites, while for the southern boreal site the most useful NDIs included reflectances in selected visible and NIR wavelengths (R 2 > 0.7) (Fig. 5).Collectively for data from all sites, the variation in the fractional cover of vascular plants could explain more than 80% of the variation in specific wavelength combinations of NDI in the spectral regions of NIR-shoulder and red (e.g., R 2 = 0.80 for λ 1 = 750 nm and λ 2 = 700 nm) (Fig. A.1).In other words, the difference between the red and NIR reflectances increased linearly as a function of increasing vascular plant cover, probably due to overall densification of understory canopy structure.
For the non-vascular plant cover, overall, the strength of the relationship with NDI was collectively weak (e.g., R 2 = 0.28 for λ 1 = 630 nm and λ 2 = 438 nm) but varied strongly between the sites.While selected NDIs for the temperate mountain and hemiboreal sites had very strong or strong relationships with non-vascular plant cover (R 2 > 0.9 and R 2 > 0.7, respectively), for southern boreal site the relationships were generally weak.R 2 was not calculated for the temperate floodplain site due to low number of stands with moss.An increase of fractional cover of non-vascular plants may have increased the reflectance of the forest floor particularly in the shortest measured wavelengths (Vogelmann and Moss, 1993).
The best NDIs for predicting intact litter cover for all sites collectively included reflectances in wavelength pairs of red (685 nm) and green to red (600-640 nm), or red edge to NIR shoulder (735-740 nm) and NIR (1180-1285 nm) (e.g., R 2 = 0.61 for λ 1 = 1230 nm and λ 2 = 740 nm).For study sites individually, NDIs with strong relationships with intact litter cover (R 2 > 0.70) were abundant for hemiboreal and temperate mountain sites across the wavelength regions, and scarce for southern boreal and temperate floodplain sites.For the latter two, the NDIs with highest R 2 s (> 0.7) included reflectances in narrow spectral bands either in the visible or NIR regions.The increase of intact plant litter cover of the forest floor probably induced both a decrease in the spectral absorptance in the red wavelengths, as well as a rounding of the NIR-shoulder, while the reflectance in the longer edge of NIR remained more of less unchanged (Nagler et al., 2000).
Overall, our analysis confirmed and generalized to several biomes the earlier (local) findings that fractional cover of different living and non-living materials on the forest floor are connected to the spectral properties of the forest floor (Forsström et al., 2021).The fractional cover values depended on tree canopy characteristics (Section 3.1 in our study; Majasalmi et al. 2020), which supports the hypothesis that also direct links between tree canopy characteristics and forest floor spectral properties may exist.This is examined in Section 3.3.Fig. 5. Coefficients of determination (R 2 , in colors) for all possible spectral combinations (λ 1 , λ 2 ) (350-2315 nm) in narrowband normalized difference index (NDI) and three forest floor fractional cover classes (in columns) present in all study sites (in rows).NDI is calculated from the mean hemispherical-conical reflectance factors as (HCRF λ1 -HCRF λ2 ) / (HCRF λ1 + HCRF λ2 ).Only stands with fractional cover ≥ 0.01 of each class are included (i.e., R 2 was not calculated for subfigure H. due to low number of stands).The number of stands (n) is given in the top left-hand corner of each subfigure.Spectral data are excluded in wavelengths with high atmospheric noise.P.R. Forsström et al.

Links between light availability and spectral properties of forest floor
We examined the variation of forest floor reflectance factors (y) in nine selected Sentinel-2 MSI spectral bands as a function of light availability (x) (DIFN) (Fig. 6).Strongest linear relationships were observed in the SWIR bands collectively in all sites (1614 nm: R 2 = 0.29 for y = -0.20x+ 0.36; 2202 nm: R 2 = 0.34 for y = -0.18x+ 0.22).This can be explained by the increase of fractional cover of vascular plants in more open canopies (Fig. 2A) and green vegetation's (incl.vascular and nonvascular plants) strong spectral absorptance in the SWIR region (Gates et al., 1965;Harris et al., 2005;Hovi et al., 2017;Forsström et al., 2019).Also, weak and decreasing linear trends were observed in the ultra-blue (R 2 = 0.15), blue (R 2 = 0.09), and red (R 2 = 0.12) bands.In other spectral bands, the relationships were unclear.
When examining relationships between DIFN and forest floor's reflectance in the nine spectral bands separately for each study site (Fig. 6), the clearest trends were noted in the temperate floodplain site (max DIFN = 0.2), where the forest floor brightened linearly with the opening of the tree canopy in the NIR bands (865 nm: R 2 = 0.63 for y = 0.74 + 0.21; 945 nm: R 2 = 0.54 for y = 0.69 + 0.25), and darkened in the red (R 2 = 0.66 for y = -0.20 + 0.09) and in the longer SWIR (2202 nm: R 2 = 80% for y = -0.49x+ 0.27) band.In the southern boreal site, the relationships between spectral properties of forest floor and light availability were mostly scattered, probably due to high level of natural variability in the site fertility (Forsström et al., 2021) which induced variation in the forest floor composition between the stands (Table 1).For all sites collectively, the strength of the relationship between forest floors' spectral reflectance and DIFN was different for stands with different dominant tree types.In mixed type stands, the relationship was strong in all bands (R 2 was always between 0.54 and 0.80; e.g., R 2 = 0.80 for y = -0.04+ 0.05 in the blue band) except in the NIR bands (R 2 between 0.14 and 0.18).In contrast, stands dominated by broadleaf trees showed moderate to weak relationships in NIR bands (e.g., in 865 nm: R 2 = 0.38 for y = 0.40 + 0.28) and very weak in the green and red bands (R 2 = 0.22 for y = 0.04x + 0.06; R 2 = 0.20 for y = -0.06x+ 0.07, respectively).For the pine dominated stands, the relationships were non-existing (R 2 between 0 and 0.01 in all bands), and for spruce stands they were unclear due to low number of stands with open tree canopies.
The links between light availability (DIFN) and spectral properties of forest floor were further investigated using R 2 of NDIs.(Fig. 7).In general, the strength of the relationship depended on the combination of wavelengths and varied notably between the study sites.While the collectively useful NDIs for all sites were scarce (e.g., R 2 > 0.35 for 1240 nm and 1220 nm, and R 2 > 0.40 for 680 nm and 670 nm), for individual  Bands 1-8, 8a, 9, 11, and 12).P.R. Forsström et al. sites the strength of best NDIs ranged from very strong in the temperate floodplain (e.g., R 2 = 0.98 for 1520 nm and 545 nm) and mountain site (e.g., R 2 = 0.95 for 1740 nm and 1670 nm) as well as in the hemiboreal site (e.g., R 2 = 0.94 for 2110 nm and 600 nm) to strong in the southern boreal site (e.g., R 2 = 0.67 for 1735 nm and 1640 nm).Overall, the southern boreal site exhibited the lowest, and the temperate floodplain sites the largest number of NDIs with strong relationship with DIFN.The best wavelength combinations (highest R 2 ) were typically observed close to the spectral absorption features of chlorophyll and water, and the variation of R 2 s between the sites can be explained by the variation in the tree types and fractional covers (Figs.2C, 3).Namely, low diversity in tree types was observed to promote a monotonous forest floor composition (Figs.2A, C) and increase the number of NDIs with strong relationship with DIFN.
The range of forest floor REPs for all stands was from 689 to 725 nm (Fig. 8), and the hemiboreal stand had REPs in longer wavelengths (710-725 nm) compared to the stands in other sites with similar DIFNs (Fig. 8A), probably due to the larger fractional cover of vascular plants (Horler et al., 1983;Curran et al., 1990;Clevers et al., 2001) (Fig. 2C).The observed wavelength range was in line with previously reported REPs of boreal (Rautiainen et al., 2011;Forsström et al., 2019) and hemiboreal (Nikopensius et al., 2015) forests during peak-growing season.However, when plotted against DIFN the data were scattered and the REPs were only weakly linked with the variation of tree canopy structure.For all stands, REPs (y) shifted towards longer wavelengths as a either a linear or exponential function of increasing tree canopy DIFN (x) (R 2 = 0.14 for y = 11.5 + 706.3;R 2 = 0.21 for y = 710 •.005x -9.5 − 12.9x ).The temperate floodplain exhibited stronger linear relationships between REP and DIFN (R 2 = 0.51 for y = 74.7x+ 698.8) compared to temperate mountain (R 2 = 0.25 for y = 8.2x + 702.9), southern boreal (R 2 ~ 0) and hemiboreal sites (R 2 = 0.24 for y = 8.7x + 713.7).
We also combined all stands belonging to the same tree type from all sites, and compared the relationships between REP (y) and DIFN (x) (Fig. 8A).Broadleaf stands exhibited the strongest relationship with REPs moving fairly linearly into longer wavelengths as a function of increasing DIFN (R 2 = 0.43 for y = 31.63x+ 704.9).Very weak linear relationships were observed also for spruce (R 2 = 0.22 for y = 12.96x + 704.8) and mixed (R 2 = 0.19 for y = 5.66 + 707.8) type stands.No such relationship was observed for pine dominated stands (R 2 ~ 0).The linear relationship observed for broadleaf stands between REP and DIFN was linked to our previously shown increase of vascular plant cover as a function of increasing DIFN (Fig. 3A) which influenced the shape of red edge curve by enhancing the absorptance in the red wavelengths and reflectance in NIR (e.g., Horler et al. 1983) in stands with more open tree canopies (Fig. 6).
The changes of forest floor REP were further investigated for effect of the variation of fractional covers of vascular and non-vascular plants, and intact plant litter (Figs.8B -D, B.1, Table B.1).For all sites collectively, the relationships were linear so that the REPs moved into longer wavelengths with increasing vascular plant cover (R 2 = 0.60 for y = 15.91x+ 702.4,n = 64), and into shorter wavelengths with the increase of intact litter cover (R 2 = 0.37 for y= -15.4x + 715, n = 67) and non-vascular plant cover (R 2 = 0.18 for y= -13.76x + 711.9, n = 50).For the study sites individually, the relationship was moderate or strong between REP and both vascular plant and intact litter cover (R 2 s between 0.55 and 0.89, and between 0.48 and 0.75, respectively) except for the southern boreal site which exhibited very weak relationship with litter cover (R 2 = 0.11).As a function of the non-vascular cover, REPs were mostly scattered and the trends varied between the study sites: a strong linearly increasing relationship was found for the temperate mountain site with small number of stands with moss cover (R 2 = 0.80 for y = 49.9 + 702.2, n = 5), a moderate linearly decreasing relationship Fig. 7. Coefficients of determination (R 2 ) for all possible spectral combinations (λ 1 , λ 2 ) (350-2315 nm) of narrowband normalized difference index (NDI) and the light availability at the forest floor (DIFN).NDI is calculated from the mean HCRFs as (HCRF λ1 -HCRF λ2 ) / (HCRF λ1 + HCRF λ2 ).In the legends are the number of stands (n) dominated by pine, spruce, broadleaf, or mixed trees, respectively.Spectral data are excluded in wavelengths with high atmospheric noise.
for the southern boreal site (R 2 = 0.32 for y = -12.92x+ 711, n = 33), and a very weak relationship for the hemiboreal site (R 2 = 0.14, n = 10).For the temperate floodplain site, the relationship between REP and non-vascular cover was not calculated due to low number of stands with moss (n = 2).
In most cases, the relationships between the shape features of the red edge, i.e., peak amplitude (dr re ) and area (Σdr 680-740 nm ), and fractional covers of vascular and non-vascular plants, and intact plant litter were similar or weaker compared to those observed with REPs (Figs.B.2, B.3).However, for the temperate floodplain site both dr re and Σdr 680-740 nm (y) resulted in nearly perfect relationships with the variation of the vascular cover (x) (dr re : R 2 = 0.99 for y = 0.0067x + 0.0006; Σdr 680-740 nm : R 2 = 0.99 for y = 0.30x + 0.03).Similarly, in the southern boreal site both shape features exhibited stronger linear relationships with litter cover compared to REP (dr re : R 2 = 0.39 for y = -0.005x+ 0.007; Σdr 680-780 nm : R 2 = 0.37 for y = -0.22x+ 0.30).Our results imply that the spectral features of the red edge are valuable descriptors of the composition of forest floor, and that the sensitivities of these different features to light availability (i.e., tree canopy DIFN) and fractional covers of forest floor vary between forest biomes.

Conclusions and future perspectives
In the beginning, we hypothesized that tree canopy structure, and hence light availability at forest floor (DIFN), is connected not only to the composition of forest floor (as has been reported in previous studies, e.g., Cole et al. 2017;Rautiainen and Majasalmi 2020) but also to its spectral reflectance properties.Based on our new findings, these properties are interconnected in European boreal, hemiboreal and temperate forests.We showed that the forest floors' spectral properties were different between our four study sites in the shortwave spectral range (350-2500 nm), and that the observed variation of the absorption and reflectance features was linked to the variation fractional covers of vascular and non-vascular plants, intact and decomposed plant litter, and lichen.Further, forest floors in stands with open tree canopies (i.e., larger DIFN) had large fractional cover of vascular plants and less plant litter, and exhibited spectral features commonly associated with green vegetation, including REP.We also demonstrated that spectral data beyond visible and NIR wavelengths are useful in differentiating forest floors in different forest biomes and have potential for estimating forest water condition.Examination of vegetation indices of forest floor revealed that NDIs in specific wavelength combinations were linked to tree canopy DIFN and could be used to predict the variation of different fractional cover classes.However, the strength of the relationship between NDIs, and DIFN and fractional cover classes varied between the study sites.
We suggest that interpretation of remote sensing data from satellites, aircrafts, and from close-range remote sensing systems (Liang et al., 2022) such as drones and mobile spectral cameras would benefit from incorporating information on the biome-specific relationships between tree canopy and forest floor spectral properties in different types of forests.For example, in operational remote sensing, the relationships can be used as constraints or input information when inverting reflectance models to estimate biophysical forest variables (e.g., Varvia et al. 2018;Schraik et al. 2019;Pisek et al. 2021) or possibly, in the future, to link routine forest inventory data sets to satellite data through allometric or biophysical relationships and forest reflectance models.In addition, multipurpose, rapid close-range spectral sensors may even be used to

Fig. 1 .
Fig. 1.Location of the study sites in Finland (southern boreal site), Estonia (hemiboreal site) and the Czech Republic (temperate mountain and floodplain sites).Photos show examples of the tree canopy and forest floor composition.

Fig. 2 .
Fig. 2. Tree canopy structures and forest floor compositions for all study sites: A. Diffuse non-interceptance (DIFN) of the tree canopy.B. Effective plant area index (PAI e ) of the tree canopy.C. Mean fractional covers of forest floor.The colors denote the four study sites, and in A. and B. the bars are standard deviations, filled symbols are mean values, and empty symbols are for individual study stands.The shapes of the symbols show dominant tree type in each stand.

Fig. 3 .
Fig. 3. Relationships of forest floor fractional cover and light availability at the forest floor (DIFN) for those fractional cover classes that were present in all study sites.Only stands with fractional cover ≥ 0.01 of each class are included.

Fig. 8 .
Fig. 8. A. Forest floors' red edge positions of each study stand in relation to light availability (DIFN).B -D. Forest floor red edge positions of each study stand in relation to three forest floor fractional cover classes present in all study sites.Only stands with fractional cover ≥ 0.01 of each class are included (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.).

Table 1
• , 15-30 • , 30-45 • , 45-60 • , and 60-73 • .The exact pixel positions corresponding to the zenithal annuli were derived from a lens function obtained from separate calibration measurements.Canopy gap fractions were then calculated per annuli as the number of sky pixel Tree canopy and forest floor layer characteristics for each study site.The values are means for study sites, and standard deviations (in parenthesis) and data ranges [in brackets] of stand level measurements.Abbreviations are for diffuse non-interceptance (DIFN) and effective plant area index (PAI e ).