Introduction 

Accurate and reliable measures of forest canopy are crucial to a wide range of studies including hydrology, carbon and nutrient cycling, and global change ([13], [41]). For this reason, forest canopy properties are widely used in many long-term research programs and, on the other hand, to monitoring forest ecosystems’ status ([15], [37]). In addition, the availability of observations on forest canopy properties such as leaf area index and forest light conditions are essential to calibrate remotely-sensed information based on airborne and satellite data ([47], [58]).

However, direct measurements of forest canopy are particularly challenging to obtain, owing to inherent difficulties in making direct measurements of forest, high levels of spatial and temporal variability, and difficulty of generalizing local measurements to the landscape scale ([12], [13], [40]).

Harvesting of trees for direct measurements in forest is labor intensive, destructive, time and money consuming, and can not be applied to large areas ([13]). Alternative and less destructive methods based on tree allometry and litterfall have been developed in order to measure forest canopy properties. Nevertheless, even these methods are labor intensive, time-consuming and not error-free because of their site- and species-dependency ([6]). By contrast, remotely sensed vegetation indexes have novel potential but still need cross calibration by means of ground-based observations ([58]). As a consequence, indirect measures of forest canopy properties using ground-based instruments have long been implemented, as documented by the rich literature (i.e., [13], [16], [6], [29], [40]).

Indirect methods enable estimation of forest canopy properties by measurements of the radiation transmission through the canopy, making use of the radiative transfer theory ([49]). However, indirect methods, including the LAI-2000 plant canopy analyzer (Li-Cor, Lincoln, Nebraska) and AccuPAR Ceptometer (Decagon Devices, Pullman, WA), two of the most commonly used devices, are hindered by the complexity of forest canopy architecture ([13]) and the high-cost of the instruments ([41]).

Since the first approach provided by Evans & Coombe ([20]), hemispherical photography (also known as fish-eye photography) has long been used for the indirect optical measurement of forest canopy. However, because of significant obstacles involving film cameras (i.e., lack of software, time-consuming acquisition and processing procedures), film hemispherical photography has been progressively forsaken ([6]). Advances in digital photographic technology have led to renewal of interest in photographic methods for indirectly quantifying forest canopy. So far, hemispherical photography is the widely used of several photographic techniques. Fish-eye photography enables characterization of forest canopy by means of photographs taken looking upward (or, in some cases, looking downward) through an extreme wide-angle lens ([29]). The method has many advantages over the other indirect methods. It is rapid, inexpensive and readily available; hemispherical image provides a permanent record of the geometry distribution of gap fraction, which is generally used to calculate forest light regimes and canopy properties such as canopy openness, leaf area index, leaf angle distribution. Hence, hemispherical photography can greatly expands the number of the canopy properties that are possible to estimate, as compared with the other indirect methods.

In spite of the recent improvement in digital photography, significant obstacles to the adoption of digital hemispherical photography still remain; accurate and meaningful estimates of forest canopy properties with digital hemispherical photography are hindered by different critical steps, regarding image acquisition and software processing; thus, adequate field collection and image processing procedure is required to achieve the standard of an ideal device ([29]).

The purpose of this contribution is to briefly introduce some of the major drawbacks of the digital hemispherical photography method. Given that different controversies of digital hemispherical photography have been usually treated separately, this contribution is aimed to: (i) provide a basic foreground of digital hemispherical photography, in order to outline strengths and weaknesses of the method; (ii) to give an update framework of the main procedure recently proposed to overcome the technical problems of digital hemispherical photography; (iii) to provide an reliable field measurement and images processing protocol for canopy description and analysis, particularly regarding sampling strategy.

Foreground to Digital Hemispherical Photography 

The first hemispherical lens was developed by Hill ([24]), to study cloud formation. The first approach to fish-eye photography in forestry was then provided by Evans & Coombe ([20]), which used hemispherical photography to describe the light environment under forest canopy. Anderson ([1], [2]) used fish-eye photography to calculate the direct and scattered components of solar radiation from visible sky directions. Subsequently, film hemispherical photography has been used for a long time to estimate forest canopy properties ([5], [3], [8], [57]). However, technical and theoretical obstacles involving many time consuming steps have progressively prevented the wide spread adoption of film hemispherical photography ([6], [41]).

More recently, advances in digital photographic technology and image processing software have led to a renewal of interest in digital hemispherical photography for indirect quantification of forest canopy properties ([6], [41], [28]). Digital cameras have greatly simplified the process of image capture and processing, when compared with film cameras ([37]). In addition, over the last few years, numerous commercial software packages, as well as freeware programs for canopy analysis, have been developed ([21], [30], [27]). Recent studies confirmed the accuracy of digital hemispherical photography in estimating forest canopy properties ([19], [30], [34], [39], [50]). Moreover, new photographic techniques have been tested recently, confirming the high potentiality of digital photography ([40], [51], [14]).

Theoretical background 

Hemispherical photography is a method that measures the gap fraction at multiple zenith angles. Gap fraction is computed by applying the Beer-Lambert law (eqn. 1):

(1)

where LAI is the leaf area index, θ is the zenith angle of view, G(θ) is named G-function and corresponds to the fraction of foliage projected on the plane normal to the zenith direction. In theory, by measuring the gap fraction at multiple zenith angles it is possible to simultaneously determine both LAI and the foliage angle distribution function ([40]). Forest light environment was also derived from gap fraction.

The method makes the following assumptions:

• Leaves are randomly distributed within the canopy;
• Individual leaf size is small compared with the canopy and thereby with the sensor field of view;
• Foliage is black, namely it do not transmit light;
• Leaves are azimuthally randomly oriented.

Current controversies and corrective strategies 

Photographic exposure

Photographic exposure affects the magnitude of the canopy gap fraction ([61]). The importance of exposure control is well documented, since automatic exposure has been demonstrated to prevent accurate and reliable estimates of the gap fraction ([11], [38], [61]). Images taken with automatic exposure underestimates gap fraction in open canopies, while overestimates gap fraction in medium-high density canopies ([61]); as a consequence, exposure needs to be manually set.

The optimum exposure for hemispherical photography would be the one which makes the sky appear as white as possible, providing in the meantime the best contrast between canopy and sky ([11]). Adequate exposure can be approximately determined in two steps:

• reference exposure is measured in a clearing (open sky), with aperture set to provide adequate depth of field;
• subsequently, exposure is set to overexpose image (generally by 1-3 stops of the shutter speed) relative to the open sky reference ([38], [61], [41]), with the aperture unchanged. This exposure setting makes the sky appears white, providing satisfactory contrast between canopy and sky, and is not influenced by the stand density ([61]).

Gamma function

Unlike film cameras, image sensors in digital cameras have the advantage of respond linearly to light ([61]). However, in order to simulate the non-linear behavior of the human eye, the in-camera software applies a logarithmic transformation by means of gamma function ([7]). The gamma function describes the relation between actual light intensity during photography and the resulting brightness value of the pixel ([54]). A gamma value of 1.0 denotes an image that accurately reproduces actual light intensity ([41]). Digital cameras typically have gamma values between 2.0 - 2.5. The main effect of this correction is to lighten the midtones, thus resulting in worse estimate of canopy light transmittance ([7]).

Some studies found that gamma correction strongly affects forest canopy properties estimates in both film and digital cameras ([54], [34], [41]); consequently, back-correcting to 1 the gamma function of the images is recommended.

Pixel classification

The optimal light intensity (brightness value) from a digital color or grey levels image is generally used as threshold value to distinguish pixels belonging to sky or canopy, thus producing a binary image ([54], [29], [30], [7]). Some authors suggested using the blue channel instead of the grey levels of the RGB image, because the foliage elements have a much lower reflectivity and transmittance in the blue region of the visible electromagnetic spectrum ([34], [61], [37]).

Previously, pixel classification was performed manually. Some software package such as GLA (Gap Light Analyzer - [21]) still employs interactive manual thresholding. However, some studies pointed out that manual thresholding could be a relevant source of error due to its subjectivity ([47], [29], [7], [28]). As a consequence, different automatic, objective, operator-independent thresholding methods have been proposed to replace manual thresholding ([26]. [30], [44], [37]), while commercial software packages (i.e., Winscanopy) typically have developed automatic pixel classification algorithms ([37]).

A detailed analysis of the different classification methods falls out of the scope of this contribution (for a more complete description, see [55], [37]). However, Macfarlane et al. ([41]) noted that correcting images for the camera’s gamma function and correcting the gap fraction distribution for foliage clumping are more important on leaf area index derived from digital hemispherical photography than the classification method chosen. In addition, Macfarlane ([37]) even found that none of the more complicated classification methods available for image processing (also from remotely-sensed imagery classification procedures) yielded results that greatly differed from a simple global binary threshold classification.

Leaf area index estimates

Three main sources of discrepancy are commonly recognized when digital hemispherical photography is used to estimate forest leaf area index.

(a) Digital hemispherical photography estimates a plant area index, rather than actual leaf area index, due to the contribution of woody elements ([6]). Deciduous forests allow the estimation of woody area index from optical sensors, which can be estimated from gap fraction during leafless ([16], [33]). For evergreen broadleaved and coniferous species, the woody material could be estimated from destructive sampling ([33]), or from tabled woody to total area ratio ([13]).

(b) Another source of discrepancy is the clumping of foliage ([6], [41]). Foliage clumping (Ω) strongly affect the canopy gap fraction, according to the Beer-Lambert law (eqn. 2, as modified by [43]):

(2)

To overcome the limit of a non-random distribution of foliage within the canopy, some commercial software for hemispherical image analysis (i.e., Winscanopy) calculates clumping indexes from an analysis of the gap size distribution ([12]) or from the gap fraction distribution of a number of azimuth segments for each annulus of the hemisphere ([31], [53]), or by combining these two approaches ([34]).

Van Gardingen et al. ([53]) demonstrated that correcting for foliage clumping can reduce the underestimation of up to 15%, compared with conventional analysis of hemispherical photography, which can results in an underestimate of 50% of the leaf area index derived from harvesting. Another advantage of fish-eye photography is that the instrument enables assessment of both within and between crowns clumping effects, which results in greater accuracy in LAI retrieval in dense canopies, when compared with LAI-2000 PCA ([14]).

(c) Even though an apparent advantage of fisheye photography is that LAI and the extinction coefficient (k) are simultaneously estimated [G-function is related to extinction coefficient by G(θ) = k · cos(θ)], previous studies found that the foliage angle distribution calculated from hemispherical photography appeared sensitive to canopy structure ([10], [39]). As such, the foliage angle distribution calculated from fish-eye images should be treated with caution. To overcome this limit, an alternative is measuring the gap fraction at a single zenith angle of θ = 57.5°, given that the extinction coefficient at this angle was largely independent of the foliage angle distribution ([4]). Some software packages allow the 57.5 degree analysis of fish-eye images (i.e., Winscanopy and CAN-EYE).

Protocol for image acquisition and hemispherical software image analysis 

In order to provide clear and concise suggestions to get the most out of using digital cameras for forest canopy properties estimation, an hypothetical application of digital hemispherical photography is illustrated, with an example of the compact camera Nikon CoolPix 4500, equipped with the FC-E8 fish-eye lens converter, and the Winscanopy software. Camera setup and software analysis was set according to Macfarlane et al. ([41]). The reason for choosing a compact camera is motivated mainly because the Nikon CoolPix models have been very popular in this field, and the performance of these cameras, as well as other compact camera models, have been deeply investigated. For instance, Frazer et al. ([22]) compared film photography with the 2.1 Megapixel Coolpix 950, Inoue et al. ([25]) compared the effect of quality and image size in two different Coolpix models (990 vs. 900), Leblanc et al. ([34]) used both Coolpix 990 and 5000 in boreal forests, Englund et al. ([19]) tested the effect of image quality using the Coolpix 950. These researchers found that little or no differences exists between TIFF and JPEG images from the same camera, but that image size can influence canopy properties estimates.

Recently, DSLR (Digital Single Lens Reflex) cameras have become much more affordable and their resolution has greatly increased ([46]), but thorough appraisals using DSLR cameras are still poorly documented; hence, generalization over canopy measurement procedures using DSLR cameras can not be achieved so far. However, we refers to the work of Pekin & Macfarlane ([46]) for a detailed analysis of the effect of quality, image size, file format, ISO in both Coolpix 4500 and DSLR Nikon D80.

Sampling strategy

Sampling strategy is a key issue when performing ground measurements that need to be representative of the whole canopy ([59]). Number of images and spatial location of shots define the sampling strategy. Canopy and vegetation type, spatial variability, plot area, sensor angle of view and distance to the edge of the stand can greatly influence the accuracy of sampling design ([9]).

It is best to consider a sampling protocol designed for the canopy type which is being measured. Canopy height is the first factor which should be considered. As a rule-of- thumb, the distance between the sensor and the nearest leaf should be at least four times the width of the leaf. As a consequence, the use of upward pointing fish-eye images in short canopies such as grassland and agricultural crops should be carefully evaluated ([33]). The distance between the lens and the canopy may be too short, and the resulting canopy covered by the field of view of the camera may be not representative of the spatial distribution of the canopy ([36]), When this situation occurs, the use of downward looking camera orientation comes as a reliable and practical alternative for agricultural crops and grassland ([18], [23], [36]). Downward pointing camera can also be used to separate understory vegetation and top canopy vegetation in a forest stand.

Canopy spatial variability is a major factor affecting sampling strategies. For closed and randomly distributed canopies, a grid of sample points is usually a suitable strategy ([32]), even though predetermined sample location may require several adjustments, in that the presence of leaves immediately above the sensor may block the entire view at low zenith angles. By contrast, Leblanc et al. ([35]) proposed the sampling along a 70 m transects over boreal and temperate forests, with measurements every 10 m.

In the case of regular tree distributions, e.g., plantations of tree in evenly spaced rows, the adoption of a crisscross array scheme is recommended to ensure sampling under trees, thus avoiding bias from inter-row gaps sampling ([13]); the sample distance should be proportional to the range of distances between rows ([59]).

Accurate samplings in open and heterogeneous canopies are more challenging to obtain. Gap fraction is greatly influenced by clumping, especially in heterogeneous canopies (see eqn. 2). Moreover, clumping occurs at different scales, from shoot level (within crown) to stand level (between crowns). This multiscale nature makes it hard to quantify foliage clumping ([50]).

Irrespective of the method used to estimate gap fraction, in most applications gap fraction is given only in term of zenith angle, since an assumption of azimuthal symmetry is generally used ([53], [33]). This assumption implies that such standard techniques should be limited to homogeneous canopies. It is well known that conifer needles are not randomly arranged in space, and radiation penetration models assuming homogeneous canopy will underestimate the transmittance of a conifer canopy. Hemispherical photography enables assessment of both within and between crowns clumping (for more details, see the section “Leaf area index estimates”. As such, the incorporation of clumping is strongly recommended, when available from software outputs (Tab. 1).

Again, heterogeneous canopies require more repetitions (images) than homogeneous canopies to achieve good spatial sampling. Image-processing software also allows to mask some part of the hemisphere, in order to reduce the field of view, which may improve spatial representation in heterogeneous canopies, i.e., to include dense and sparse regions of a heterogeneous canopy in separate images.

The masking procedure could also be used in mixed forests, in order to sample clusters of different species in different images. Masking can also be used to prevent some undesired part of the image from being analyzed (i.e., sun glint, operator, etc).

As previously outlined, use of downward pointing camera enables analysis of understory, and even allows separating this component from top of canopy elements. In the case of taller understory, Rich et al. ([48]) suggested using tall-folding monopod with self-leveling mount set-up to sample top of canopy, or, alternatively, using a ladder. Use of a ladder also enable measuring canopy at different heights, which could be useful in tropical wet forests.

Other sampling difficulties arise from measurement on single trees, because indirect methods are poorly suitable for single plants ([17]). However, Hemiview software proposes specific options for measurements on single trees ([48]). The LAI-2000 user’s manual proposes similar suggestions, which can also be suited to hemispherical photography.

Specific precautions should be adopted for slope, such as holding the lens normal to the ground; e.g., self-leveling tripod is provided with Winscanopy equipment. Some authors even suggested corrective methods to introduce slope effect in the analysis ([56], [52]).

Image acquisition

Hemispherical images should be collected in summer, under fully developed canopy conditions, and under uniform overcast sky, or alternatively close to sunrise or sunset ([34]); both these sky conditions enable a perfect diffuse sky, thus avoiding the interference of direct sunlight, which can cause errors of up to 50% ([60]).

Images should be collected as fine quality and at maximum resolution JPEG, with the lens set to F1, which enables circular images. Lens set to F2 enables full-frame fish-eye image instead of circular image, with the former having a better resolution than the circular format. However, only recent releases of Winscanopy software (since 2006a version) have implemented analysis of fullframe image; so far, canopy analysis has been usually performed only on circular images ([40]). Camera must be aligned to magnetic north and pointed upward by means of a self-leveling tripod. The aperture must be set to minimum (5.3) and, with the camera in aperture-priority (A) mode, the exposure must be recorded in an adjacent clearing. Subsequently, the mode must be changed to manual (M) and the shutter speed must be lowered by two stops in comparison to the exposure metered in the clearing ([61]). Exposure should be measured regularly beneath the canopy using a spot light meter, in order to check possible changes in sky conditions during image acquisition ([41]).

Different exposures can be promptly collected by setting exposure bracketing, which automatically adjust the shutter speed from the starting exposure, which is set by the operator (i.e., the open sky reference). On the other hand, digital cameras which can save image files in RAW format, such as DSLR allows varying the exposure after image acquisition.

Software image analysis

Gamma function of the images needs back-correction to 1 prior to hemispherical software image analysis. Given that Nikon CoolPix 4500 has a gamma function of approximately 2.2 ([34]), the original images must be adjusted with the gamma correction set to 0.45 (1/2.2), using a standard image manipulation program such as Irfanview ([40]).

In the blue band of the electromagnetic spectrum, the foliage appears darker than in the other bands, thus minimizing the interference of multiple scattering in the canopy and chromatic aberration ([61]). In addition, in diffuse sky conditions, the sky is saturated in the blue band, and thus appears white in 8-bit blue channel ([33]). As such, the blue channel of the images should be used in the canopy analysis to achieve optimal brightness value (thresholding). Image must be sharpened (medium), to enhance the contrast between sky and canopy, and then analyzed using hemispherical image analysis software. Winscanopy enables automatic pixel classification of the images, thus avoiding human input.

In addition, Wincanopy enables correction for clumping foliage, which can significantly reduce the underestimation of leaf area index in clumped canopies ([31], [53], [29], [14]).

A zenithal angle range of 0-70° and 8 azimuth segments should be adequate for the image analysis ([41]).

Comparison of software packages

The more popular commercial software packages are Winscanopy and Hemiview. Their standard systems include a digital camera, a calibrated fish-eye lens and a self-leveling tripod. Free software packages are available for hemispherical image analysis such as GLA (Gap Light Analyzer - [21]) and CAN-EYE.

Most of the scientific studies concerning hemispherical photography use method based on the determination of optimal threshold (Hemiview, GLA, Winscanopy). Moreover, most of these studies focused on forest canopies ([18]). CAN-EYE is also widely used in agricultural environments, because of its ability to perform different pixel classification procedures, as compared with thresholding method, thereby allowing analysis of downward-looking images ([18]). Tab. 1 lists the main characteristics of some of the most widely used software packages.

Conclusive considerations 

Despite uncertainties due to image acquisition and processing steps, digital photography holds great promise for estimating forest canopy properties, on account of its speed, ready availability and low-cost, which enables widespread use of the method. Photography even shows good potential to replace other indirect methods, due to its ability to provide simultaneously several parameters characterizing solar radiation and forest canopy properties ([13]). In addition, unlike other methods, hemispherical photograph can be interpreted as a map of canopy openings (or, on the contrary, of canopy closure) relative to the locations from which image is taken, which can be inspected to provide insight into heterogeneity within a canopy and to compare different canopies at different sites ([48]). Last, but not least, digital photography enables widespread use of the method. Aside from scientific purposes, photography can be suitably applied for management and monitoring issues, i.e., routine canopy properties estimation.

Recent advances in digital photographic equipments such as higher resolution cameras and better quality lenses, combined with robust and efficient image processing routines and software packages, are bringing digital photography to a mature stage, where the field techniques and image processing steps are no longer significant obstacle limiting its application ([37]).

Acknowledgements 

This research was supported by RI.SELV.ITALIA Research Program 3.1 - “Silviculture, productivity and conservation of forest ecosystems” research project and by the Research Program D.M. 19477/7301/08 - “Maintenance of collections, databases, and other activities of public interest” funded by the Italian Ministry of Agriculture and Forest Policies.

References