forest ecosystems

Marc Bouvier1*, Sylvie Durrieu1, Frédéric Gosselin²; Basile Herpigny²

1

UMR TETIS Irstea-Cirad-AgroParisTech/ENGREF, Maison de la Télédétection en Languedoc-Roussillon, 500, rue J.F. Breton BP 5095, 34196 Montpellier Cedex 05, France

2

UR EFNO, Irstea, Domaine des Barres, F-45290 Nogent-sur-Vernisson, France

Submitted to Remote Sensing of Environment, ForestSat 2014 special issue

2.1 Abstract

We explored the potential of lidar for modeling the link between floristic biodiversity and forest structure under a Bayesian framework. Understory vegetation was studied in terms of plant species abundance and richness of ecological groups based on light preference. Lidar data provided the opportunity to access relevant indicators related to spatial arrangement of the vegetation. Variables derived from lidar data were found to improve Bayesian models linking biodiversity indicators to environmental variables. Several lidar variables measured at the plot level or in a wider neighborhood around the plot were found to have both significant and non-negligible relationships with biodiversity indicators. The responses of biodiversity indicators to forest structure variables were shown to be dependent on environmental conditions characterizing the study area. Shade- tolerant and heliophilous species richness were impacted by indictors linked to light penetration through the vegetation cover. The study also highlighted that forest structure in the neighborhood of field plots can impact biodiversity indicators measured at plot level. Various scales were found relevant to describe forest structure in abundance and richness models. Finally, results highlighted the interest of using lidar data to characterize forest structure in different ways, with variables that can hardly be measured in the field, either due to their nature or to the size of the area they describe.

Keywords: Biodiversity, Airborne lidar, Plant diversity, Plant richness, Forest structure

Forest ecosystems cover nearly one third of the global land area and contain over 80% of the world terrestrial biodiversity (FAO, 2011). Biodiversity conservation is imperative in the face of increasing anthropic pressures and threats to forest ecosystems (ONU, 1992). Our ability to evaluate and monitor biological diversity is essential to ensure effective conservation. Forest structure is a key factor driving several processes in forest ecosystems. Stand structure affects the microclimate and habitat quality and therefore biodiversity potential. Several studies have highlighted the existence of a link between forest structure and certain biodiversity indicators, including wildlife richness (MacArthur and MacArthur, 1961; Carey et al., 1991) and floristic diversity (Brunet et al., 1996; Dale, 1999). Understory vegetation is known to be particularly sensitive to forest structure (Getzin et al., 2012). Strong evidence for inter-dependency between understory vegetation and forest structure has been found by relating different forest structure indicators to spatial disturbance structures (Gilliam and Roberts, 1995; Whigham, 2004; Barbier et al., 2008). However, what drives plant species distribution and composition over different forest habitats is still unclear (Tuomisto et al., 2003).

Forest structure usually refers to the spatial arrangement of the vegetation. Structure may be characterized by tree height, height heterogeneity, a horizontal canopy and gap distribution. Variables related to the vertical distribution of vegetation in a stands are known to be of primary importance (Durrieu et al., in press) and dominant height is widely used in ecological modeling to define site fertility (Bontemps and Bouriaud, 2014). Simonson et al. (2012) highlighted the importance of variables related to mean canopy height for plant species composition and diversity. Several variables of canopy height heterogeneity were found relevant when predicting hazel grouse (Bonasa bonasia) occurrence (Zellweger et al., 2014). Gaps have a short-term impact on biodiversity through different mechanisms. They result in an increased irradiance that can benefit heliophilous species and also have an impact on the microclimate in the patch, with a higher temperature variance in gaps than under the canopy. Vascular plant development can be also positively impacted by the removal of trees, which results in an increased release of water and nutrients (Durrieu et al., in press). Bouget (2005) reported that the gap effect was favorable on the whole to saproxylic beetle biodiversity, with the notable exception of shade-preferring groups. Benefits have also been observed for forest vascular plants; Duguid and Ashton (2013) found that silvicultural systems based on small gaps favoured floristic biodiversity. Furthermore, gaps do not only have an effect in the open area they

This is particularly true among birds (Germaine et al., 1997).

Establishing models that are able to reliably describe the link between biodiversity and forest structure remains challenging; however, such models would facilitate the implementation of sustainable management strategies and practices. Two main avenues of research could be explored to make advances in this field. The first one is to develop modeling frameworks adapted to both the complexity and the specificities of the link between biodiversity and environmental conditions. The second one is to develop methods that will improve the capacity to measure and describe three-dimensional vegetation structures, in order to have appropriate and reliable structure indicators as input variables for biodiversity models.

The Bayesian statistical approach has recently become popular in ecological research (Ellison, 1996; Clark, 2005) and is a highly promising framework to tackle issues of biodiversity modeling. The Bayesian approach makes it possible to draw inferences on large numbers of variables and parameters describing complex relationships intrinsic to ecological studies (Clark, 2005). As with other parametric statistical methods, Bayesian statistical models provide an estimate of the magnitude of the relationship between biodiversity indicators and ecological variables. They offer many advantages, among which a great flexibility of use which allows the researcher to integrate new probability distributions to characterize ecological count data and to shift to nonlinear models. One of the main limitations of Bayesian models – shared with other parametric statistical methods –lies in the large quantity of calibration data needed to precisely estimate model parameters when the model contains several explanatory variables (Harrell, 2001). Zilliox and Gosselin (2014) studied the link between floristic biodiversity and both abiotic and biotic environmental variables following a Bayesian approach. The authors found that forest structure variables estimated from field measurements had a non-negligible relationship with species richness for selected floristic ecological groups, an effect that varied among ecological groups and ecological conditions.

Forest structure is generally described from field measurements in biodiversity studies. Only a limited number of plots are inventoried because the field work is both costly and time consuming. Furthermore, some structural indicators relevant for biodiversity studies are difficult to assess using ground surveys (Müller and Brandl, 2009; Zellweger et al., 2014). Remote sensing has the potential to provide quick accurate structure measurements over large areas, including for metrics which are difficult or very

data make it possible to assess forest structure variables at various scales. Getzin et al. (2012) showed that gap distribution is a major driver of understory plant diversity in deciduous forests and demonstrated the use of high-resolution aerial images in surveying understory species. However, optical sensors do not provide the three-dimensional (3D) information needed to characterize forest structures (Wulder, 1998; Hall et al., 2005; Richardson and Moskal, 2011). To this end, the potential of lidar (Light detection and ranging) systems has been widely acknowledged (Dubayah et al., 2000; Lim et al., 2003b). Lidar are active remote sensing systems based on the emission and reception of laser pulses. They provide precise distance measurements between the equipment and a target by measuring the time elapsed between pulse emission and the reception of the echo reflected by the target. The spatial positions on the Earth’s surface of the targets that caused significant backscattered echoes are calculated from distance measurements and from the position and orientation of the lidar system. They are measured with a differential global positioning system (DGPS) and an accurate inertial unit. Airborne Laser Scanners (ALS) combine a lidar and a scanning system. ALS systems can therefore record data over a swath, the width of which depends on both scanning angle and flight altitude. Apart from a few full-waveform systems, most ALS systems record data in the form of a point cloud. In forest environments, where light can reach the ground through gaps, the incident pulses will be backscattered by the top of the canopy, the understory vegetation and the ground. Thus the resulting 3D point clouds embed information on the 3D structure of the vegetation and are processed to assess diverse lidar variables describing forest structure. The use of lidar in landscape ecology research and ecology studies is quite recent. Lidar technology provides an opportunity to build more variables describing different aspects of forest structure from those observed or measured during field surveys. These new variables might be more suitable for describing the link between forest structure and understory vegetation. Variables extracted from lidar data have been proposed to characterize landscape patterns and structures (Mücke et al., 2010; Vierling et al., 2008). Some studies have also used lidar to map understorey vegetation (Hill and Broughton, 2009; Martinuzzi et al., 2009).

Regarding biodiversity, lidar data have been used to analyse relationships between certain biodiversity indicators and a broad range of structural variables related to the three- dimensional arrangement of vegetation (Bradbury et al., 2005; Lesak et al., 2011; Müller and Brandl, 2009; Zellweger et al., 2013, 2014; Müller et al., 2014). The first studies investigated the relationships between lidar variables and faunal biodiversity. Floristic biodiversity has been explored only more recently and to a lesser extent. Simonson et al.

oak forests. They found a lidar-measured vegetation height positively associated with species diversity (R2=0.50). Lopatin et al. (2014) also used lidar data on Mediterranean forests in order to predict plant richness from an object-based model (R2=0.64). In this study, topographic variables were shown to be more relevant than the selected lidar variables to predicting plant richness. Indeed, the ecological context should be systematically integrated, in addition to vegetation structure variables, when models explaining biodiversity indicators are being built. Ecological context refers to all the variables, on which biodiversity indicators highly depend (Maestre et al., 2009). Floristic biodiversity is controlled by both abiotic environmental variables (i.e. regional climate, soil parent material, topography and the time) and biotic factors (Major, 1951; Austin and Van Niel, 2011). Complementing lidar variables with environmental variables should improve the predictive power of floristic biodiversity models.

The aim of this study was to evaluate the potential of lidar variables to explain variations in floristic biodiversity through a Bayesian modeling framework. Two specific objectives were identified: (1) to better understand the link between understorey biodiversity - assessed through the species abundance and richness of three ecological plant groups - and certain characteristics resulting from the spatial arrangement of the canopy as assessed through lidar data ; (2) to determine if local floristic biodiversity can be influenced by surrounding structure, and not only by the immediate structure at the floristic survey point, by analyzing the impact of forest structure indicators on the studied biodiversity indicators at several scales.

2.3 Materials

2.3.1 Study sites

Two forested areas in North-eastern France were studied. A lowland forest comprised of multi-layered deciduous stands (Lowland site), and a mountain forest comprised of coniferous, deciduous and mixed stands (Mountain site). The Lowland site is a 10,000 km² area located in the Lorraine region (48.53° N, 5.37° E). The regional climate is semi-continental and subject to an oceanic influence (Joly et al., 2010). The Lorraine lowland forest is fragmented and intensively managed. In the selected area, forests are dominated by European beech (Fagus sylvatica L.), European hornbeam (Carpinus betulus L.) and Sycamore maple (Acer pseudoplatanus L.). The Mountain site is a 9,340 km² area located in the Vosges region (48.03° N, 7.08° E). The regional climate is semi-continental

from about 120 m to 1420 m. The stands are typically heterogeneous and uneven-aged, and are dominated by European beech, European silver fir (Abies alba Mill.) and Norway spruce (Picea abies (L.) H.Karst).

2.3.2 Lidar data

Data were collected at both sites using small-footprint airborne lidar systems. For each site, lidar data only covered a sub area. For the Lowland site, a 60 km² sub-area was surveyed in October 2010 at a flight altitude of 550 m a.s.l. by means of an LMS-Q560 (Riegl, Austria). The sensor operates at a wavelength of 1550 nm. The scan angle was set to 29.5° and the final pulse density was 20.7 pulses/m². For the Mountain site, a 1,200 km² sub-area was surveyed in March and April 2011. An ALTM 3100 lidar system (Optech, Canada) was used at flight altitude of 1500 m a.s.l.. The sensor operates at a wavelength of 1064 nm. The scan angle was set to 16° and the final pulse density was 3.4 pulses/m².

Data pre-processing was performed for each study area by the data providers, Sintegra (France) and French National Institute of Geographic and Forest Information (IGN, France), for the Riegl and Optech data respectively. Ground points were classified following the TIN-iterative algorithm (Axelsson, 2000) in order to produce a digital terrain model (DTM). Then, first return points were extracted from the data to produce a digital surface model (DSM). Both the DTM and DSM had a 1 m resolution. For each acquisition, aboveground heights were calculated by subtracting from each lidar point elevation the corresponding ground elevation given by the DTM, thereby removing topographic effects from the lidar point clouds. From the resulting lidar point clouds, four sub-point clouds were extracted for each field plot; these point subsets corresponded to various spatial extents around the study plots where local biodiversity was assessed in the field.

2.3.3 Field plot data

Data were collected in 789 circular plots within the Lowland site. For forty-eight plots (9 m radius) within the sub-area covered by lidar data, field data were obtained from the EcoPlant database (Gégout et al., 2005) and, for the 741 plots (15 m radius) outside the sub-area, data came from the IGN. At the Mountain site, data were obtained from the IGN for 1,155 circular plots, 171 (15 m radius) of which were surveyed within the sub-area covered by lidar data and 984 (15 m radius) of which were outside the sub-area. In both study sites, field plot data were collected from 2008 to 2012. Field plots covered in ice and snow during the inventories were excluded from the dataset. Field plot center positions

differential corrections to improve position accuracy, which was assumed to vary between 0 and 4 meters. Field plot data include information on floristic biodiversity, several abiotic environmental factors and one biotic factor related to local forest structure.

2.3.4 Information on floristic biodiversity

Understory species were identified and their abundance estimated at each field plot. A Braun-Blanquet cover class (Braun-Blanquet, 1964) was attributed to each of eight species in each study region (Table III.1). Species richness was also estimated for ecological groups based on light preference. Three species classes were distinguished based on Ellenberg values 1 to 9: shade-tolerant, from 1 to 3 (shad); intermediate-light, from 4 to 6 (mid); and heliophilous species, from 7 to 9 (helio) (Ellenberg et al., 1992). Julve's (1998) autecological table of correspondence was used to assess the Ellenberg values of each species.

Table III.1. List of the eight most abundant species in each study sites. Species are ranked in order of decreasing abundance.

Lowland site Mountain site

Species name Species code

Ellenberg

value Species name

Species code

Ellenber g value

Brachypodium sylvaticum

(Huds.) P.Beauv. brsy 4 Carex pilulifera L. capi 5

Carex sylvatica Huds. casy 5 Deschampsia flexuosa

L. defl 8

Galium odoratum (L.) Scop gaod 3 Hedera helix L. hehe 3

Hedera helix L. hehe 3 Oxalis acetosella L. oxac 4

Lamium galeobdolon (L.) L. laga 4 Rubus idaeus L. ruid 5

Milium effusum L. mief 5 Vaccinium myrtillus L. vamy 5

Anemone nemorosa L. anne 4 Digitalis purpurea L. dipu 5

Poa nemoralis L. pone 7 Athyrium filix-femina

Soil characteristics, i.e soil pH (Reaction) and soil water capacity (SWC), were derived from the mean the Ellenberg values of the understory species at both sites. Temperatures (Tmean; °C) and global solar radiation (Solrad; MJ/m²/day) were based on

May to September average values. Tmean was obtained from the French national

meteorological service (Météo-France). Solrad was estimated from temperature data using the Allen (1997) equation. Topography was described by three variables, i.e. Elevation, Slope, and Aspect. Aspect was defined as the magnetic azimuth (grades) of the plot’s largest slope. For further details on the estimation of these variables see appendix A in Zilliox and Gosselin (2014). Table III.2 provides descriptive statistics that summarize the abiotic factors for each study site.

Table III.2. Summary of field variables for both study sites.

Field variables

Lowland site Mountain site Mean Sd Min Max Mean Sd Min Max

Reaction 5.2 0.6 3.5 6.7 4.1 0.5 2.9 6.3 SWC 5.1 0.3 4.5 6.9 5.2 0.3 4.4 7.7 Tmean (°C) 9.3 0.4 8.7 10.5 8.6 0.7 6.2 10.2 Solrad (MJ/m²/day) 639.3 22.5 563.1 682.5 614.9 39.8 479.3 710.1 Elevation (m) 300.7 79.5 108.0 488.0 524.7 224.0 120.3 1,419.8 Slope (%) 9.6 7.4 0.0 29.4 27.1 18.8 0.0 85.0 Aspect (grades) 208.3 123.2 0.0 400.0 210.6 114.0 0.0 400.0 Ctot (%) 76.7 29.7 0.0 150.0 83.5 24.1 2.5 170.0

2.3.6 Biotic environmental factor

The total tree crown cover (Ctot; %), was estimated at both study sites from field

inventories (Table 2). Ctot is the ratio between the projected surface area of all individual

In this study, we focused on floristic abundance and richness modeling. We considered abundance for the eight most representative species on each study site, excluding woody species. We also took into account species richness for three ecological groups based on light preference, i.e. shad, mid, and helio. We developed Bayesian models to link species richness and abundance with both environmental and lidar variables. As lidar data did not cover the whole area for both sites, we used a two-step approach to build the models. The first step aimed at calibrating the model to define the parameters for the abiotic environmental variables for all the plots in both sites, and the explanatory biotic variables collected in the field, i.e. cover rate Ctot. The second step aimed at assessing the

potential of the explanatory biotic variable computed from lidar 3D point clouds to explain local floristic biodiversity. Several lidar variables were identified to be used as descriptors of the 3D vegetation structure in floristic statistical models. The variables were extracted from circular plots with the same radius as the field plots (9 m at the Lowland site and 15 m at the Mountain site), but also for other scales with 50 m, 100 m and 200 m radiuses. The lidar variables were tested individually to assess both the magnitude and direction of the relationship between biodiversity indicators and the forest structure components characterizing the plot itself or its wider surrounding environment. Data processing was performed in the R statistical environment version 3.1.1 (http://www.r-project.org/).

2.4.1 Lidar variables

A set of variables was identified to describe the 3D spatial arrangement of the canopy (Table III.3). Maximum height (Hmax), median height (Hmedian) and mean height (Hmean) estimated from lidar return heights were related to vegetation height characteristics.

Variation in canopy height (𝜎_𝐻2), the Gini coefficient (Gini) and the coefficient of variation for leaf area density (CvLAD) were used to measure the dispersion of lidar return heights and

inform on tree height heterogeneity within a forest stand. Maximum gap size (Gapmax),

cover fraction (Cf) and cover rate (Cr) are related to the horizontal canopy and gap

distributions. In addition, total canopy volume (Volcan) was used to capture both the

vertical and horizontal organization of the vegetation material. Vegetation points below 2 m were considered to belong to the understory and were not taken into account as tree vegetation points when we computed some of the variables related to forest structure. Therefore, the variables Hmean, 𝜎𝐻2, Gini, CvLAD, Gapmax, Cf, Cr variables were assessed with

Table III.3. Description of forest structurevariables derived from lidar data, with zi corresponding to the aboveground height of a lidar point i, n to the total number of lidar points and N to the total number of 1 m2 grid cells in the plot.

Lidar variable Description

Hmax Maximum point height: Hmax = max(zi) Hmedian

Median point height (all points, including ground points):

Hmedian = median(zi)

Hmean Mean point height above 2 m: Hmean =

1 𝑛⁡∑ 𝑧𝑖

𝑛 1

𝜎_𝐻2 Variance of point height above 2 m: 𝜎_𝐻2 = ⁡1 𝑛⁡∑ (𝑧𝑖− 𝐻𝑚𝑒𝑎𝑛) 2 𝑛 1 Gini

Gini coefficient above 2 m: 𝐺𝑖𝑛𝑖 = ⁡∑ (2𝑖−𝑛−1)⁡𝑧𝑛1 𝑖

∑ 𝑧𝑛1 𝑖(𝑛−1)

(Lexerød and Eid, 2006)

Gini has a theoretical minimum value of zero, expressing perfect

equality when all lidar points are of the same height value; it takes a theoretical maximum value of one, indicating greater diversity when all lidar points except one have a height value of zero.

CvLAD

The coefficient of variation in leaf area density above 2 m was calculated as the ratio of the standard deviation to the mean of the leaf