Procedures for predicting habitat and structural ...

3 downloads 0 Views 2MB Size Report
Apr 15, 2013 - aerial photography, videography imagery is collected as a series of frames which ...... conifier forests using hemispherical photography and fine.
Australian Forestry

ISSN: 0004-9158 (Print) 2325-6087 (Online) Journal homepage: http://www.tandfonline.com/loi/tfor20

Procedures for predicting habitat and structural attributes in eucalypt forests using high spatial resolution remotely sensed imagery N. Coops , D. Culvenor , R. Preston & P. Catling To cite this article: N. Coops , D. Culvenor , R. Preston & P. Catling (1998) Procedures for predicting habitat and structural attributes in eucalypt forests using high spatial resolution remotely sensed imagery, Australian Forestry, 61:4, 244-252, DOI: 10.1080/00049158.1998.10674747 To link to this article: http://dx.doi.org/10.1080/00049158.1998.10674747

Published online: 15 Apr 2013.

Submit your article to this journal

Article views: 18

View related articles

Citing articles: 6 View citing articles

Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=tfor20 Download by: [The University of British Columbia]

Date: 02 February 2017, At: 13:16

244

Procedures for predicting habitat and structural attributes in eucalypt forests

Procedures for predicting habitat and structural attributes in eucalypt forests using high spatial resolution remotely sensed imagery. N. Coops 1, D. Culvenor2, R. Preston 3 and P. Catling4 'CSIRO Forestry and Forest Products, Private Bag 10, Clayton South VIC. 3169. 2 Department of Forestry, Usiversity of Melbourne, Parkville VIC. 3052 3 State Forests of NSW, Locked Bag 23, Pennant Hills, NSW 2120 4 CSIRO Wildlife and Ecology, PO Box 84, Lyneham, ACT 2602 Revised manuscript received 16 June 1998

Summary

such as growth plots to a wider, spatial, context.

Forest resource information is increasingly needed at fine spatial scales for operational and strategic applications including monitoring indicators of ecologically sustainable forest management, planning of harvesting operations and implementation of silvicultural prescriptions, and the maintenance of biodiversity and ecological sustainability. High resolution remotely sensed imagery is one data source that can provide cost effective information for forest management. This paper presents two methodologies that allow these data to be modelled to predict forest structure in eucalypt forests. One method emphasises the tree crown as the primary indicator of forest structure and utilises algorithms which automatically delineate tree canopies in high spatial resolution data. The second method investigates the spectral variability of the forest in relation to its habitat quality (biomass of tree canopy, shrubs, ground cover and litter) for ground-dwelling fauna. Both methods utilised the near infrared (NIR) region of the electromagnetic spectrum as it provides an indication of photosynthetic activity of the canopy and describes most spectral and spatial variation in the scene. Application of these methodologies indicated significant relationships between forest structure (measured as either habitat quality or canopy characteristics) and the NIR variance. In the case of automated tree crown delineation there were good relationships between the variation within the delineated crowns and field estimated canopy characteristics. A significant relationship was also found between the habitat complexity scores estimated at each plot and the maximum level of variance in the near infrared channel at each plot. These preliminary results indicate that high spatial resolution imagery can be used to predict forest structure and allow predictions to be spatially extrapolated. Ultimately the ecologically sustainable management of Australia's forest resource will depend on the availability of high resolution predictions ofthe diversity and extent of forest structure and habitat over large areas of high productivity forests.

Currently, forest structure mapping is being undertaken at widely varying spatial scales. At a fine scale, forest growth and yield models tend to be based on height, diameter and basal area measurements, whereas broad-scale descriptions will typically rely on field observations and their relationship to broad structural classes such as forest communities or types. A problem with these approaches is the scale-specific nature of the structural descriptions (O'Hara et al. 1995), making it difficult to make direct comparisons between inventories at different scales. One possible solution is to sample the whole forest, that is, adopt a total inventory approach, effectively providing a detailed base layer from which structure can be described at successively broader scales, thus providing an approach compatible across a variety of management levels. The use of airborne video systems as a remote sensing tool, has greatly expanded over the past decade (Manzer and Cooper 1982; Escobar et al. 1983; Everitt et al. 1991; Mausel et al. 1992; King 1995) and is receiving increased application in resource management. Videography and other systems such as digital cameras have a number of advantages over conventional aerial photography including its cost efficiency, ready availability of imagery for visual assessment and digital processing, digital recording format, and the ability to collect data in a number of spectral wave-bands. Like aerial photography, videography imagery is collected as a series of frames which may require mosaicing if large areas are to be monitored, and radiometric corrections are required to remove distortions caused by the camera system and reflectance effects (King 1995). Other aerial imaging systems, including the Compact Airborne Imaging Spectrometer (CASI) (Harron et al. 1992), digital frame cameras such as the Kodak DCS (King 1995), and digital ortho-photography, are also useful sources of remote sensing imagery. A number of high resolution satellite sensors with quoted resolutions of down to 0.8 mare also scheduled to be launched in late 1998 and 1999. As introduced by Preston (1997b), digital videography is one of the more promising sensors for application to Australian forestry, but it is not possible at this point to reach conclusive decisions as to which particular sensor or combination of sensors are most suitable for operational forestry applications.

Introduction The need for accurate, timely, and cost-effective forest information has never been more critical. Australia's commitments to sustainable forest management and conservation of biodiversity require detailed knowledge of the existing forest resource, and a greater understanding of the dynamic processes of forest growth and changes in stand condition. A stand structure approach to forest management therefore presents a unifying theme for multiple resource management (O'Hara et al. 1995), providing information for (i) sound silvicultural practice (Florence 1996), (ii) assessing suitability of the forest for wildlife, (iii) assessment of fire risk, (iv) stratification of the forest enabling more efficient ground-based inventories and (v) facilitating the extrapolation of point source data

Applications of aerial video imaging cameras have ranged from shoreline oil spill surveys (Owens and Robilliard 1980) to imaging of soil (Vlcek 1983) and assessing rangeland conditions (Pickup et al. 1995). Forestry applications of videographic data have included the quantification of plant leaf radiation (Gausman et al. 1983), prediction of size, density and spatial distribution of leaves, branches and trunks (Wu 1988), and deriving the pattern of forest destruction related to arsenic concentrations (Airo1a 1989). Additional re-

Australian Forestry Vol 61, No. 4 pp. 244-252

search has included the prediction of soil colour and organic matter content (Mausel et al. 1990), detecting weed infestations on rangelands (Everitt et al. 1990), using middle infrared data to detect differences in forest species (Everitt et al. 1986), and detecting the presence of wildfires (Everitt et al. 1989). In Australian forestry conditions, Coops and Catling ( 1997) described how airborne video data, if correctly preprocessed, could be used to accurately predict the suitability of the forest for ground-dwelling fauna using habitat quality ratings, and from these predictions develop maps of habitat quality across the landscape. It was concluded that these spatial predictions could then be used to stratify the landscape into regions for predicting the distribution and abundance of some faunal groups. These predictions were verified in the field with species abundances being accurately predicted in over 70% of situations (Catling and Coops 1998). In related research, Preston (1997a) developed a monitoring scheme for predicting forest species and structure from airborne videography data and expert spatial models. The scheme utilises 2 m spatial resolution videographic data as well as high spatial resolution layers from a Geographic Information System (GIS) including radiation and climatic layers. This research has also been expanded to look at a broader range of forest structural and compositional attributes (Preston et al. 1998). This paper examines the use of airborne videography as a means of providing data on tree and stand structural attributes in eucalypt forests, and to facilitate the spatial extrapolation of these attributes across the forested landscape. Once these methods have been developed they may provide consistent base data for the compilation of structural information across a range of scales. The ability of forest managers to develop policies to protect and/or manage forest structure in native forests will depend on the availability of maps that accurately indicate the diversity and extent of structure over the landscape. Because forest structure can change over a few hectares, field based collection of information over thousands of hectares is both expensive and logistically impractical. Additionally, temporal changes in structure create a complex spatial mosaic as the forest undergoes progression from regrowth to senescence as well as being affected by disturbances such as fire and timber harvesting. This dynamic mosaic requires forest structure to not only be identified and quantified across the landscape, but also to be regularly monitored (Gill and Nicholls 1989). Two methods are being developed to extract forest structural information from airborne multi-spectral high resolution imagery. 1. One method emphasises the tree crown as the primary indicator of forest structure. This is convenient, as the crown is the only portion of the tree clearly visible to above-canopy sensors. However it is also consistent with a general trend toward 3-D descriptions of canopy architecture (Curtin 1970; Zeide and Pfeifer 1991; Fournier et al. 1996; O'Hara et al. 1995) as a means of understanding forest structure and associated dynamic processes. 2. The second investigates the spectral variability of the forest in relation to its habitat quality (biomass of tree canopy, shrubs, ground cover and litter) for ground-dwelling fauna. The imagery provides an indication of photosynthetic activity of the plot and is thus sensitive to all components within the plot including canopy, understorey and ground vegetation and litter, soil and shadow. This provides a highly variable response which is related to the suitabil-

245

ity of the site to fauna habitat.

Methods The study area The study area was located within the Eucalyptus dominated Currowan State Forest, Batemans Bay region of south-eastern Australia (35° 35'S, 150° 07'E) (Figure 1). Within the study area 29 sites of 0.25 hectares were established using local field and I :25,000 scale maps. Sites were placed close to features such as roads to aid in the location of these sites on the imagery. Each site was located within a homogeneous forest type and was chosen using a 'gradsect' approach allowing altitude, lithology and vegetation types to be represented (Austin and Heyligers 1991).

Figure 1. Location map of South Coast of N.S.W. Field data collection Forest stand data was collected using the Bitterlich's or relaskope method. This technique is based on a point frequency principle and defines a circular plot around which the trees are sampled, the size of which varies according to the number of trees counted. At each site tallied, trees that fell within the plot allowed calculation of basal area in m2ha· 1 (Grosenbaugh 1952). Plot locations were determined using differential GPS and additional site information was recorded including slope, aspect, local horizons and disturbance history (as assessed by an experienced field ecologist). Percentage cover was visually estimated at each plot for five features ( 1) canopy cover; (2) shrub cover; (3) ground vegetation cover; (4) the amount of litter, fallen logs and rocks and (5) a moisture rating. A score was given for the complexity of ground-dwelling fauna habitat (Newsome and Catling 1979) on the basis of the percentage cover estimates. Each percentage estimate was rated on a scale of 0-3 and the scores of the five features tallied to a total following the methodology of Newsome and Catling (1979). Generally a score offourorfive denotes a forest with poor structure with few understorey shrubs and little ground cover, whereas a score of nine or 10 typifies a forest with a thick understorey, and good ground and litter cover. ~ score of seven denotes a forest with a moderate structure.

Airborne video data The video system utilised was a Digital Multispectral Video

246

Procedures for predicting habitat and structural attributes in eucalypt forests

System (DMSV) consisting of a synchronised 2x2 Cohu 4910 series CCD camera array. Each camera has 752 (horizontal) x 582 (vertical) active picture elements and is equipped with a 12 ffim lens. The system has four bandwidth interference filters centred on 450 nm, 550 nm, 650 nm and 770 nm. The image is captured and recorded using a 32 bit Vista framegrabber mounted in a 66 Mhz 80486 personal computer. Using a 12 mm lens the DMSV system allows up to 20% overlap between frames with pixels as small as 18 em assuming an aircraft forward speed of 40 to 50 m s- 1 and an altitude above ground level of 250 m. The complete specifications of the camera are detailed in Pickup et at. ( 1995). The system was flown in a twin engined Cessna Titan flying at an altitude of 3300 m above ground level and at a ground speed of 185 km h- 1 resulting in a ground spatial resolution of 2m. Three videography transects were flown, each 20 km in a north-south direction on the 2nd Aprill996 under conditions of strong, uninterrupted sunlight with solar zeniths between 40.3° and 40.6°. The raw data from the video system were acquired and processed by the Centre for Arid Zone Research (CSIRO Division of Wildlife and Ecology) following the methods of Pickup et at. (1995) involving geometric and brightness correction and normalisation of individual scenes within each transect. Geometric correction was required to remove the effects of aircraft motion which vary with the imaging and recording system (Pickup et at. 1995). Brightness corrections were required to remove distortions caused by lens shading and vignetting, atmospheric attenuation and bi-directional variations in the ground target reflectance (Pickup et at. 1995; King 1995). In order to preserve the image variance the imagery was not geometrically rectified to a map base.

warping of the local variance surface with the highest variance occurring along the gully with either side having lower local variances.

Figure 2(a). Wire-frame representation of reflectance from tree crown (green band) subset from original video data

Processing methodology Without some kind of spatial structure, remotely sensed imagery, be it acquired from a plane or satellite, would resemble random noise (Strahler et al. 1986). Increasing spatial resolution usually results in improved mapping accuracy by facilitating a greater visual recognition of spatial features, rather than simply providing more spectral information. A number of techniques exist to measure the spatial properties of remotely sensed images including correlograms, fourier transforms and semivariograms (Curran 1988). Woodcock and Strahler ( 1987) found that where the spatial resolution of the image is considerably smaller than the objects in a scene (in this case trees), the neighboring pixels will be highly correlated and thus the variation would be small. If the objects in the scene approximate the size of the pixel then the likelihood of the neighbourhood resembling the central pixel decreases and the local variance increases. As the pixel size increases, and many objects are found within each pixel, the local variation again stabilises or decreases. The spatial nature of the high resolution imagery is shown in Figure 2. Figure 2(a) demonstrates the spatial variation associated with a medium sized eucalypt tree crown. The peak reflectance at or near the centre of the crown, and the spectral minima near the edges illustrates a relationship between the geometric and radiometric crown shape. Figure 2(b) demonstrates the spatial nature of the imagery at a plot scale with the variation of the imagery changing as the density, size and species of the forest vegetation varies over the landscape. In this example the centre of the image contained temperate rainforest found in a gully environment whilst the vegetation on either side was typical of a drier aspect. This is shown by the

Figure 2(b), Wire-frame representation of variance from a 120 x 120m forest site (NIR band) subset from original video data The processing methodology of the two approaches will be described separately.

Derivation of habitat quality To derive habitat quality indices from the videography we modified the local variance method originally proposed by Woodcock and Strahler ( 1987). These authors computed the mean of the standard deviation from a 3x3 moving window on successively spatially degraded images to compute optimal spatial resolutions (Townshend and Justice 1988; Marceau et at. 1994). In this modified technique we compute the mean value of the standard deviation using successively larger windows of pixels ( i.e. 3 x 3, 5 x 5 .... 49 x 49). Each of the 29

Australian Forestry Vol61, No.4 pp. 244-252

247

sites were located on. the videography and the local variance computed on the assumption that changes in local variance provided an indication of forest structure complexity, and thus the habitat quality of the site for ground dwelling fauna. A local variance graph for each pixel on the image provides an indication of the spatial nature of the elements within the scene. As with the Woodcock and Strahler ( 1987) method, variation occurring within a stand of forest was low at the initial spatial resolution where the window size was significantly smaller than the size of tree canopies in the scene. As the window size increased and approximated the object size (around 20 m) the modified local variance increased then peaked. However, unlike the Woodcock and Strahler ( 1987) method, the spatial resolution of the image is not being successively degraded, so the local variance does not fall substantially, but is maintained in much the same way as a semivariogram (Curran 1988).

Accuracy assessment data

This modified local variance was computed for all four spectral channels individually. The maximum level of local variance reached, for each spectral channel (independent of the window size at which it occurred), was then compared to the habitat quality score at the site.

Results

Delineation of individual tree crowns High resolution images of a forested environment may be likened to a mountainous landscape (Gougeon 1995) with the tops of trees appear brightest, being directly illuminated by the sun, while the gaps between trees are darker due primarily to directional reflectance and shading from adjacent crowns. The spectral maxima (peaks) and minima (valleys) are the primary image features used for the identification of crowns, being indicative of crown centroids and boundaries respectively. The Tree Identification and Delineation Algorithm (TIDA), described in this study, delineates individual trees based on a prior understanding of the interaction between crown structure, associated scene elements (such as exposed soil and understorey vegetation) and incident sunlight. Initially the algorithm distinguishes between vegetation and non-vegetation pixels using user-specified thresholds. Using the near infrared region of the spectrum, TIDA then performs a fourway linear search (horizontally, vertically, and in both 45 degree planes) which locates the spectral peaks within the image. The number of times that a single pixel is identified as a spectral peak provides a probability of the given pixel belonging to the spectral peak of a tree crown. Spectral minima are also located using a four-way linear search, however, unlike the maxima, pixels are identified as minima regardless of the number of positive searchers. A series of filters are then applied to refine the identified minima pixels into a clear and continuous network of absolute boundaries. Once the spectral maxima and the crown boundaries have been identified each crown is then delineated individually. The delineation process is commenced by examining the properties of pixels spatially adjacent to a computed crown centroid. Using user--specified spectral and distance thresholds a pixel is either accepted or rejected as a member of the crown using a 'top-down' approach beginning from the centroid and finishing either at minima boundaries, or user-supplied spectral boundaries, whichever is reached first. Output statistics were generated for each of the tree crowns identified, including mean reflectance and variance within each defined crown, and spatial characteristics such as area, boundary irregularity and distance to neighbouring crowns.

In order to assess the accuracy of the predictions a second dataset was developed. The most appropriate method to examine the accuracy of the TIDA predictions is to establish a database of individual tree characteristics including accurate measurements of tree crown size, stem diameter, growth stage, height, location and species. This database could then be used to match trees in the field with respective canopy objects defined by the algorithm. Collection of these data is underway, however in the short term, a smaller dataset was established to provide an initial test of the predictions of the TIDA output. Additional field sites were established along a 1 km transect of the video data to encompass a range of forest structure. Ten sites of 0.25 ha were located along the transect and percentage cover was estimated for the five forest structure components as described previously.

Habitat Quality Prediction Of the four wave-bands collected by the videography the nearinfrared wave-band demonstrated the greatest variation (the highest local variance). This was expected as green foliage reflects strongly in the near-infrared region of the spectrum. Accordingly the difference between the canopy, understorey, soil and shadow provided a highly variable response in the video imagery (Coops and Catling 1997). Figure 3 shows an example of the modified local variance, as computed for three sites from the videography at increasingly larger window sizes, and demonstrates the levels of variance obtained over sites with different habitat quality scores. ~-------------------------------------. 35

30

~Max. Local

v-

Low-

Complexity

10

5L-----------------------~----~~--~ 0 w ~ 30 ~ ~ ~ Spatial Rooalution (m)

Figure 3. Graphs of the local variance for three selected sites at Currowan State Forest. The X axes indicates the spatial resolution with the Y axis showing the level of variance at increasing window sizes of pixels for the NIR channel (adapted from Coops and Catling 1997). The high habitat quality site (complex forest structure) was rated a score of 9 which indicated a highly complex moist forest with high percent cover of overstorey, understorey shrubs, ground cover and litter. The medium site represents a habitat quality score of 7 which indicated a dry forest of lower complexity than the most complex site with reduced understorey and ground cover. The lowest site was dry and exposed with a sparse understorey of shrubs and grouoo cover which resulted in a habitat quality for ground dwelling fauna score of 5 (Coops and Catling 1997). The relationship between the maximum local variance reached in the NIR channel of the video data and the habitat quality

248

Procedures for predicting habitat and structural attributes in eucalypt forests

scores at the 29 sites was highly significant (p < 0.001) with the majority of sites falling within the 95% confidence intervals of the linear model. A linear regression of the two variables produced a model with an R2 of0.75 (n=29) and a standard error of 0.68 habitat quality score units (Coops and Catling 1997).

Accuracy assessment of the habitat predictions Field verification has shown that 80% of the predicted habitat quality scores for ground dwelling fauna exactly matched the recorded scores at the site (Coops and Catling 1997) with the predicted scores in the remaining cases being very close to that recorded in the field.

Figure 4(a) shows the NIR image of the 1 km x 1 km subset of the videography area of Curro wan State Forest, New South Wales. Figure 4(b) shows the extrapolation of the relationship detailed above across the landscape of the I x I km area as predicted habitat quality scores.

Figure 5(a). Subset of NIR image of Currawon State Forest. Area is indicated by white box in Figure 4(a).

Figure 4(a). NIR image of the I km by 1 krn subset of the videography. White box indicates extent of region shown in Figure 5(a).

Figure 5(b). TIDA Output for the subset of videography in Figure 5(a).

TIDA predictions

Figure 4(b). Prediction of Habitat Scores across the 1 km by 1 krn subset of the videography. Habitat scores are shown from white to black with the lowest habitat score predicted (5) corresponding to white and the highest score (10) corresponding to black.

Initially, all four input bands of the videography were examined to establish which was the most appropriate waveband to use for the tree delineation process. The NIR band was selected due to its sensitivity to variation in canopy shape compared with the three visible bands and provided the strongest variation between the canopy, understorey, soil and shadow. Similar work on individual tree delineation has also concentrated on the NIR wavebands (Gougeon 1995; Fournier et al. 1995). The image shown in Figure 5(a) was processed using the TIDA methodology and statistics were computed for all the deline, ated trees on the I x I km subset of the imagery. Figure 5(b)

Australian Forestry Vol61, No.4 pp. 244-252

shows the output of the TIDA algorithm on a subset of the 1km x 1km Currawon image. TIDA produces a variety of levels of output. In Figure 5(b) the individual tree boundaries are presented for visual comparison with the original videography image. Accuracy assessment of the TIDA predictions To provide an initial accuracy assessment of the TIDA predictions, individual tree statistics were averaged for each of the ten accuracy assessment plots along the lkm transect (as described in Coops and Catling 1997) and compared with the forest structural components measured in the field using simple correlation analysis. Simple correlation analysis provides a basic method of looking for dominant trends, but may not detect other important but more complex relationships. These tabulated results are given in Table l which indicate that the field variables most significantly correlated with the TIDA output are the total habitat score (summation of all five forest structure components) and total canopy score (summation of the three canopy components only). The most significant TIDA output was the within crown variance in the NIR waveband averaged for each plot and the mean red reflectance of the crowns for each plot. The relationship between the average crown variance in the NIR waveband and the total habitat score is not unexpected given the modelled relationship between total habitat quality scores and maximum local variance in the NIR channel described earlier. This positive relationship suggests that as the forest structure becomes more complex, the mean plot variation within the individual canopy elements also increases in the NIR waveband. Negative correlation between the mean red reflectance of each of these individual canopy elements and total habitat score implies greater absorption of red light in highly structurally complex forest plots. This again is to be expected as increased photosynthetic activity which occurs in typically healthy vegetation (indicative of a structurally complex forest stand) leads to greater absorption of red radiation. Similarly, as the forest becomes less structurally complex the effect of ground cover, rocks and exposed soil

249

may become greater leading to a lightening of the soil background and subsequently an increase in the red reflectance.

Discussion Continued improvement in remote sensing technology is seeing the complementary development of image analysis and interpretation techniques. These new methods typically take greater advantage of the spatial patterns common to all images of the earth's surface, and make inferences based on a priori knowledge of the interactions between incident light and forest elements. The prediction of forest structural variables and the identification and delineation of individual trees in high resolution imagery is one example of the application of these image analysis tools. These tools can then be applied over large areas for an array of possible ecological applications. In particular it has the potential to provide fine resolution, accurate estimates offorest structure attributes, including species, spatial distributions of the trees, stem diameter, height and growth stage. It also has significant appeal as a tool to supplement traditional field based programs and provide a base for longterm monitoring, an important component of ecologicallysustainable forestry and scientific management (Gill and Nicholls 1989). As with any remote sensing techniques, airborne videography data are not devoid of spectral and spatial distortion. Hence the need to address a number of issues when applying these data to the prediction of forest structure. These issues are discussed below. Pre-processing of video data In the past, videography data has been avoided for a number of applications because of real or perceived calibration problems. These areas of concern have predominantly been with (a) geometric correction of the video data and (b) calibration of the video signal to absolute radiance or reflectances and (c) minimising the changing effects of illumination across a single frame of the image. Geometric correction is required to remove the effects of aircraft motion of which the most

Table 1: Correlation results (r) ofTIDA output with field scores within the Currawon state forest. Characteristics of Canopy Objects

Upper canopy(%)

Middle canopy(%)

Lower canopy(%)

Total canopy score(%)

Ground cover(%)

Total Habitat score

Average crown area Variance in crown area Average within-crown reflectance (red) Average within-crown reflectance (NIR)

-0.18 -0.30 -0.14

0.53 0.70* -0.68*

-0.25 -0.48 0.61

0.61 0.62 -0.72*

0.03 0.58 -0.62

0.23 0.66* -0.81**

-0.08

-0.11

-0.07

-0.34

-0.01

-0.11

Average within-crown reflectance (green) Average within-crown variance (red) Average within-crown variance (NIR) Average within-crown variance (green)

-0.16

-0.41

0.30

-0.55

-0.41

-0.52

-0.17

0.36

-0.27

0.30

-0.22

-0.11

-0.10

0.75*

-0.68*

0.62

0.76*

0.86**

-0.12

0.58

-0.50

0.49

0.05

0.21

*p

Suggest Documents