Canopy structure from below and above: relations between hemiphots and an Ikonos image Flores O. a,∗ , Tr´ebuchon J.-F. a , Gond V. a , Gourlet-Fleury S. a a CIRAD

- UPR Dynamique foresti`ere, TA 10/D, Campus international de Baillarguet, 34398 Montpellier Cedex 5, France

Abstract

Key words:

1

Introduction

In the tropics, competition for light is a critical process for forest dynamics. Under vegetation cover, light availability strongly influences trees species regeneration. In turn, light conditions depend on canopy structure which can be defined as the spatio-temporal organization of the morphological elements that compose the upper vegetation layer of a stand. Techniques to study canopy structure are usually classified as direct when measures involve physical manipulation of the elements that compose the canopy (e.g.. leaf area, leaf angle), or indirect when measures concern integrative properties of the canopy (e.g.. canopy openness, FPAR). Among indirect methods, hemispherical photographs (hemiphots) allow to quantify light conditions under forest cover. They are commonly used in studies of tree species regeneration (???). The analysis of hemiphots provides three types of variables as regards canopy properties. Canopy structure is commonly described in terms of canopy openness and Plant Area Index. From ∗ Corresponding author. Email address: [email protected] (Flores O.).

Canopy structure from below and above

the sun path, the duration of sunflecks can be calculated knowing the geographical coordinates and orientation of the photographed scene. Finally, the amount of solar energy received locally can also be estimated in terms of total, direct or diffuse radiation, given hypotheses about light diffusion in the vegetation cover and estimation of cloudiness. Hemiphots constitute a relatively easy-to-use technique to estimate this three types of variables, although it imposes constraints such as grazing light during acquisition and threshold choice during computation. A relatively newer technique, still rarely used to study tropical forests dynamics, is the remote sensing of the Earth surface which allows data acquisition at various spatial scales and resolutions. In the fields of population biology and community ecology, satelite-based images have become a relevant source of information because ecosystems can be studied at resolution of 1meter or less (??). In tropical forests, this technique served to measure stand structure (?), diameter growth of tree crowns (?) and mortality rates of emergent trees (?), or to detect skid trails, logging damage and gold panning sites (?). Most of the studies using remote-sensing in tropical forests were concerned with the geometry of various objects. In the present study, we were interested in the spectral properties of the forest canopy at the experimental site of Paracou (French Guyana). In this context, we evaluated the potentiality of Ikonos images to characterize the structure of the canopy. In particular, we asked the following question: how do spectral properties of the canopy provided by an Ikonos image relate to canopy structure as seen from below through the analysis of hemiphots ? For this purpose, we analyzed the relationships between canopy descriptors derived from the analysis of hemiphots and from the analysis of an Ikonos image.

2

Material and Methods

Study site and sampling design The study was conducted at the experimental site of Paracou (5◦ 18’ N, 52◦ 23’W) in a terra firme rain forest. The site is dedicated to the study of the impact of sylviculture on the dynamics and composition of the forest. It lies in the coastal part of French Guiana in an equatorial climate. At Paracou, streams incise a smooth geomorphological system mainly on shallow ferralitic soils. The experimental design of the site consists in 300 × 300 m permanent sample plots (PSP) arranged in four blocks. In these plots, three sylvicultural 2

Canopy structure from below and above

treatments were applied between 1986 and 1988 combining selective logging and additional poison-girdling (A detailed description of the treatments can be found in ?). For this study, the sampling design consisted in 261 points located in three PSP of the Paracou southern block, i.e. an undisturbed control plot and two treated plots (Fig.1). Variables of canopy structure from hemiphots At each sampling point, an hemiphot was taken 1.5 m above the ground with a digital camera (Nikon Coolpix 995). We analyzed the images with the imaging software GLA 2.0 (?) and computed the following variables for each: - PAI : Plant Area Index ?, CO20 and CO50 : canopy openness measured in solid angles of 20◦ and 50◦ from zenith positions (Fig.2). Spectral properties from Ikonos image We derived spectral properties of the canopy from an Ikonos image of the study site taken in July 2002 (Fig.1). Ikonos images are composed of three monochromatic bands with a resolution of 4 m and a panchromatic band with a resolution of 1 m (Table 1). The treatment of the image was conducted as follows: (1)- oversampling of monochromatic bands at 1 m-resolution, (2) positioning of the sampling points on the image using geographical coordinates, (3) elimination of points under cloud cover (n = 9), (4) pixel selection around each point: 1 m×1 m pixels completely or partially included in a 5-m radius disk were retained (Fig.1), (5) statistics calculation (mean and standard deviation) based on pixel reflectances in the four bands. The NDVI (Normalized Difference Vegetation Index ) was also calculated RN IR −RR as (?): N DV I = R , where RN IR and Rr are the reflectances in the N IR +RR near infra-red band and in the red band respectively. We then inspected relations among the calculated variables through a Principal Component Analysis. Three variables were finally retained to limit their number and avoid information redundancy across variables: - the mean reflectance in the green band: Grem , the mean NDVI : N DV Im , - the standard deviation of the reflectance in the panchromatic band: Pans . 3

Canopy structure from below and above

Statistical analysis We studied two types of relationships between the variables derived from the hemiphots and from the Ikonos image. Two points were removed from the initial set because of missing data. First, we inspected univariate relationships through Spearman rank correlation analysis, which is non-parametric and thus do not rely on hypotheses regarding variables distribution. Second, we inspected multivariate relationships between the two groups of variables through a canonical correlation analysis (CCA). Canonical correlation analysis seeks independent linear combinations of the variables in one group that are best correlated with independent linear combinations of the variables in the other group. The analysis treats the two groups of variables symmetrically (?).

3

Results and Discussion

In tropical forests, light conditions are major determinants of competition among tree species that strongly influence their regeneration. In this study, we investigated the possibility of gaining access to these conditions from remotesensing which allows to deal with large areas at fine resolution. Our approach consisted in relating the structure of the canopy, characterized by hemiphots, to spectral properties derived from an Ikonos image of a tropical rain forest. Canopy structure was described through Plant Area Index and canopy openness in two solid angles. CO20 accounted for gap fraction in a narrow angle while CO50 integrated openness in a wider angle. From 2 to 41% of the canopy were detected as gaps in the 20◦ -angle and from 4 to 21% in the 50◦ - angle (Tab.2). PAI can be considered as an indicator of the vegetation bulk and varied between 2.7 and 4.6 m−2 .m−2 . (comparaison avec autres sites?) Regarding spectral properties of the canopy, we retained three simple variables after analyzing multivariate relations a larger set. The NDVI is commonly used to detect photosynthesis activity and hence active vegetation. It varies between -1 and 1 by definition, and it is generally positively related to measures of vegetation amount. In the present case, NDVI values were between 0.4 and 0.7 (Table 2), because all sampling points were under forest cover. Distributions of Grem and P ans were less directly interpretable because these variables were designed in a heuristic purpose. The reflectance of the canopy in the green channel (Grem ) was 32% in average (Table 2). About one third on the energy reaching the canopy in this channel was reflected. In the panchromatic channel, reflectances had a standard deviation of 8% in 4

Canopy structure from below and above

average. Variability greatly differed among canopy descriptors. Regarding P AI, Grem and N DV Im , low coefficient of variation indicated similar low variability (Table 2). CO50 and P ans showed intermediate variability while CO20 showed the highest variability among the six descriptors. Hence, the three latest variables were the most likely to distinguish sites with different properties regarding canopy structure. Univariate correlations were low among variables of the two different groups, whereas the highest and most significative values appeared within the two groups (Fig.3). These features indicated relative redundancy in the information brought by the variables of a group and meanwhile little overlap between the two groups. The Canonical Correlation Analysis aimed at finding synthetic axes of variability within the two groups of variables. In each group, these axes were independent and maximized correlation with the axes obtained from the other group. Hence, the analysis found the linear relationships that were best related and thus helped to identify information common in the two groups of variables. The correlations measured between the axes defined in the two groups of the analysis were 0.19 (p < 0.01, Spearman rank correlation test) for the first axis, 0.14 (p < 0.05) for the second and 0.07% (p > 0.05) for the third axis. Only the first two axes are discussed in the following text. Among variables derived from hemiphots, canopy openness in the 20-◦ angle (C020 ) was strongly positively correlated with the first axis of this group (Fig.4). Canopy openness in the 50-◦ angle (C050 ) also showed positive, although weaker, correlation with the first analysis. PAI was weakly and negatively correlated with the first axis. This first axis thus reflected a gradient of canopy openness mostly detected in the narrow angle. C050 ) and PAI both showed positive correlation with the second axis of the analysis (Fig.??). Along this axis, sampling sites showed simultaneously higher overall canopy openness and higher vegetation bulk which thus reflected a gradient of heterogeneity in canopy structure. Clumping in vegetation components could account for the increase in the two parameters (C050 ) and PAI ). Among Ikonos variables, NDVIm strongly correlated with the first axis derived from this group. Hence the gradient of NDVI values was related to the gradient of canopy openness detected in the other group of variables. Hence, photosynthesis activity reached higher levels in the most open sites compared to mean conditions. We propose that such correlation could be explained by the photosynthesis activity of vegetation layers under the canopy. In closed sites, sub-canopy layers are shaded and not visible from above, while in more

5

Canopy structure from below and above

open sites, light energy reaches those layers and may increase photosynthesis activity relatively. Meanwhile, low canopy openness must be interpreted carefully because the measure of openness obtained from below the canopy actually integrates vegetation bulk from the measure point to the canopy top. Thus, canopy closure can arise at different, including low, heights along the vertical vegetation profile. The second axis was positively correlated to P ans and weakly negatively correlated to Grem . P ans measured the variability of relectances in the panchromatic band at a fine resolution (1 m) and over a small surface (∼ 75 m2 ). It thus reflected local heterogeneity. In the group of variables derived from hemiphots, the second axis was also interpreted as an heterogeneity axis. In the Ikonos group, the relationship between P ans and the second axis thus indicated that the highlighted heterogeneity of reflectances related to the heterogeneity of canopy structure as evidenced by hemiphots variables. The variability of reflectances at local scale could be related to clumping of canopy elements such as leaves.

4

Acknowledgements

6

Canopy structure from below and above

Tables Band

monochromatic

panchromatic

blue green red near infra-red

λ (µm)

Resolution (m)

0.45 − 0.52 0.52 − 0.60 0.63 − 0.69 0.73 − 0.90

4

0.45 − 0.90

1

Table 1 Wavelength range (λ) and spatial resolution of Ikonos bands.

7

Canopy structure from below and above Ikonos minimum mean maximum c.v.

Hemiphots

Grem

N DV Im

P ans

C020

C050

P AI

26.3 32.1 37.5 0.06

0.408 0.587 0.701 0.07

3.27 8.31 14.15 0.26

1.5 12.5 41.6 0.54

4.1 10.2 20.8 0.22

2.7 3.4 4.6 0.07

Table 2 Summary of variables distribution (n = 250). Grem : mean reflectance in the green channel (%), N DV Im : mean value of the Normalized Difference Vegetation Index (ratio), P ans : standard deviation of the reflectance in the panchromatic channel (%), C020 , C050 : canopy openness measured in 20◦ and 50◦ solid angles (%), P AI: Plant Area Index. c.v.: coefficient of variation (m2 .m−2 ).

8

Canopy structure from below and above

Figures

Fig. 1. Image of the study site (Ikonos, panchromatic band) showing the limits of the permanent sample plots and the position of the sampling points in the control plot (S-W) and in two disturbed plots (N-E and S-E). Circles figure the 5-m radius disks used for image treatment.

9

Canopy structure from below and above

Openness (%)

20

15

10

5

0 0 Z

10

20

30

40

50

60

70

Angle (°)

Fig. 2. Left : Hemiphot in negative black (canopy gaps) and white (vegetation). The image is divided in 10 × 10◦ sectors. Right : Distribution of canopy openness vs zenith angle (Z: zenith at the center of the image, graph obtained with Gap Light Analyser 2.0 ?).

10

80

90

Canopy structure from below and above

0.40

0.37 ***

+

Pans −0.15 *

0.22 ***

+ + + + + ++ ++ ++ + + ++ + + + + + + + ++ + + + +++ + +++++++ + +++ ++ +++ +++ ++++++ + ++ +++++++ + ++ +++ + + + +++++++ + +++ +++ + + ++ ++ + + + + ++ + + + + + + + + + ++ + + + + + + + + + +++ ++++ + + ++ + ++ + +++++ + + ++ + ++ + + ++ + + + + + + + + + ++ + + + + + + ++ + + + + + + + + + + + ++ ++ + + + +++ ++ + + + + + + ++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ + + + ++++++++ + ++ + ++++ +++++ + ++++ +++ +++++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ + +++++ + +++ ++ ++++ + + + +++ ++++ ++++ ++ +++ ++++ + + + + + + + + + + + + + + + + ++ + + + + + + +++++ + + + +++ + + ++ +++++ +++ +++ +++++++++ ++ ++ ++ +++ +++ + + + + + + + + + + + ++++ + + + + + + + + + + + + + + + + + ++ + + +++++++++ + ++++ +++ + +++ + ++ +++ + + + + + ++

30

CO20

0.15 *

−0.0095

0

10

−0.041

++ + ++ ++ + + ++++ + +++ + ++ ++ ++++ ++ ++ +++ + +++ ++ ++ + ++ ++ ++++ + +++++++ ++++++ ++++ + + ++ + ++++ +++++ + + + + + ++ ++ ++ + + + + + + + + + + +++++ + +++++++ + + + +++ + + + ++ +++ + + + + +++++ +

CO50 0.06

0.20 **

0.035

0.34 ***

++ ++ + + + ++ ++++ +++ + + ++ + ++ + + ++ ++ + ++ + ++ + ++ + + ++ ++ ++ ++ +++ ++ ++ ++ +++ ++ ++ + +++ ++ + + + + + + + ++ + + ++ ++ + + ++ + + + + + + + + + ++ + ++ ++ ++ ++++ ++++ + + ++ + + ++ ++ +++ +++ +++++ ++ +++++ ++ ++ +++ + + + ++ ++ ++ + + + + + + ++ +++ ++ + ++ ++ ++ + ++ + ++ + + + ++ ++ ++ + ++ + ++ + + + ++ +++ +

26

30

+

34

4.0

+

+

12

0.55

NDVI

3.0

+ + + + + + ++++ + +++ ++ ++ + +++ +++++++ ++++ + + + + + + + + + + ++ ++++++ + ++ + +++ ++ ++ + ++ ++ ++ + + +++ ++++ ++ ++++ ++ + ++ + + ++ ++ + ++ ++++ ++ ++++ +++++ ++ +++ ++ ++ ++ + ++++++++ ++ +++ + + + ++ + + + + + + + + + + ++ + ++ +++ + ++ ++++++++ ++ + +++ +++ ++ + +++++ ++ ++ + ++ ++ ++ + + + ++ + + + + + + + + +++ + + + + + + + + ++ + + + + + + + + + + ++++ ++++ +++ ++ ++ +++++ ++ + + + + + ++ + ++ ++ + ++++++ + + ++ + ++ ++ + +++++++ + + +++ + +++++ +++ +++++++++ + + +++ + + ++++ + + + ++ + + + + + ++ + + + + + + + + + + + + + + + + + + + + +++ + + ++++ + ++ + +++++ +++ +++ +++ + +++++ ++ ++ ++ +++ +++ ++ + + ++ ++ + +++ + ++ + + + + +++++++ + + + + + + + + + + + + + + + + + + + + +++++ + + + + + +++ +++++ ++++++++ + ++ ++ + ++ ++ +++++ + ++ ++ + + +++ + +++ +++ ++ ++ ++ ++ +++++ +++ + + + +++ + +++ ++++++ ++ +++ ++ ++++++ + +

4 6 8

10 20 30 40

+ 20

0

15

0.70

+

0.70

+ + ++ + + + + + + ++ ++ +++++ ++ + + +++++++++ + ++ ++++++++++ +++ ++ + ++++ + + + + + + + + + + + +++ + + + + + + ++++++++ +++ + + +++++ +++++ ++ ++ +++ + +++ ++ +++ +++ +++ +++++++ + + ++ + +++++ + ++ +++ +++ ++++++ +++++++ ++++ + +++++ ++++++++ + + + + ++++ + ++ + ++ ++ ++ ++ ++ ++ + + +++ ++++++ +++ ++ + + ++ ++ + + ++ ++ ++ ++++++ ++ + + ++ + + ++ ++++ ++ +++ + ++ ++ ++ ++ + ++ ++ ++ ++ + + + + + + + + + + + + + ++ + + + + + + + + + + + + + + + + + + + + ++ + ++++++ ++ +++ + ++++ ++++ + ++++ ++++++++ ++ ++ + + +++ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + ++ +++ ++ +++ +++++ ++ ++++ +++++ ++++++++++++++++ + ++ +++++++++ + + + + + + + + + + + + + + + + + +++++ + ++++++ ++++++ + +++ ++++ + ++++++ + + + +++ + ++++++++ + + +++++ + + ++ ++++ + + ++++ + + + ++ + + + ++ +++ + ++ +++++++++ + ++ + ++ ++ +++ ++ +++++ +++++ + ++ +++ ++ + + + ++++ ++ ++ m + + + + + + + + + + + + + + + + + + ++ + + + + + + + + + +++++++++++++++ ++ ++++ + + ++ ++++++++ ++ +++ ++ + + +++++ + ++++++ + + ++ + + + + + + + + + + + + + + + + + ++ + + ++ +++ + + ++ + + + + ++++++ + ++++ + + + + + + + ++++++++++ + + + + +++ + + + + + + + +++ + + + + ++ + + + + +++ + + + ++++++++++ + + + + + + + + + + + ++++ + + + + + + + + ++++++++++ ++ ++++ +++++++++ + +++ + + ++++ + ++ + + +++ ++++ + ++ +++++ +++ +++ + ++ + +

10

Grem

0.55

+

5

0.40

4.0

PAI −0.081

0.039

0.0099

−0.55 ***

3.0

−0.048

26

30

34

4 6 8

12

5

10

15

20

Fig. 3. Diagram showing univariate distribution of the variables (diagonal), bivariate plots (upper panel) and associated Spearman correlation coefficients (bottom panel, p-values: ∗∗∗ < 10−3 , ∗∗ < 10−2 , ∗∗ < 0.05).

11

Canopy structure from below and above

Pans

CO50 NDVIm

PAI

CO20 Grem

Fig. 4. Correlation circle showing the first two axes of the Canonical Correlation Analysis performed on the two groups of variables. From Ikonos image: Grem : mean reflectance in the green channel, NDVIm : mean value of the Normalized Difference Vegetation Index, Pans : standard deviation of the reflectance in the panchromatic channel. From hemiphots C020 , C050 : canopy openness measured in 20◦ and 50◦ solid angles, PAI: Plant Area Index.

12