and analyze them in two ways: accuracy on a known image and generalization ability on an unknown image. We show that the best choice is Gaussian distributions with full covariance matrices. We present the minimization of both terms. Our data terms can be found in section 4.2. In section 4.3 we present an exhaustive series of experiments combining the different geometric prior models with the different data terms.
than that in the visible band, due to the cellular structure in the leaves. Hence, vegetation can be identified by high NIR but generally low visible reflectance. This property was used in early reconnaissance missions in war for “camouflage detection”.
0.4 0.5 0.6 0.7 0.8 0.9 1
0 50
Wavelength (µm)
Reflectance (%)
Blue Green
Red
near IR
Figure 4.1: Plot of vegetation reflectance against wavelength.
For our experiments, we used CIR images provided by the French Forest Inventory (IFN) and panchromatic images provided by the Hungarian Central Agricultural Office, Forestry Administration (CAO, FA) taken by analogue cameras. Herein, we introduce the main properties of the images. The CIR image capturing time is usually close to 1200PM, and in most cases the exact time is available. The flight height is∼4000 m and the focal length of the camera isf = 300 mm. The film has three different spectral layers in the 520-900 nm wavelength domain. The film layers have three typical pigment colours,i.e.
cyan, magenta and yellow. Each of them is dedicated to capturing the primary colours of light. The film layers are tuned so that they extract the 900 nm, 700 nm and 520-600 nm, respectively. This technique is called the false colouring infrared representation, which translates the green into blue, the red into green and the infrared colours into red, respectively (see figure 4.2). The film has an interesting feature, namely that the second and the third layers have sensitivity almost three times higher than the first layer, to compensate for the fact that the reflectance of healthy vegetation is significantly stronger in the near-infrared domain. The panchromatic images provided by the Hungarian Central Agricultural Office, Forestry Administration (CAO, FA) are grayscale images. The capturing times were between 900AM and 1200PM, during the summer of 2002. The magnification is1 : 10000, the flight height is∼1680 m, and the camera focal length is 153 mm.
The last step before digital image processing is digitization. We need to take into ac-count the Shannon sampling theorem (Shannon, 1948), namely the sampling interval should be chosen to be less than half the size of the smallest interesting detail in the image. In our case, the images have∼0.5 m/pixel resolution. This implies that the size of the smallest details we are able to handle is 1 m. The usual size of the tree crowns that we are interested in, is between 3 m and 9 m , so this image resolution is satisfactory. In figure 4.3, we can see a typical CIR image.
Figure 4.2: False colouring technique.
Figure 4.3: Real color infrared (CIR) image with forest and urban area c°IFN.
4.1.2 Individual tree crown delineation
If we observe a forest in an aerial image, it contains trees in different positions next to each other. We can distinguish three different types of configurations, as shown in figure 4.4, for which different image features can be used by visual systems. In the first case (left), each tree stands individually. To extract the exact crown-shape it might be sufficient to threshold the image. In second case (middle), the local maximum might be useful to detect the bright tree top; contour extraction or valley detection (Eberly et al., 1994) might also be able to separate the crowns. In the third case (right), a group of overlapped trees without grey-level valleys, the detection of the individual trees must be based on the visible edges or contours.
Now we overview the most important results published on this topic. Three fundamental approaches exist to delineate individual tree crowns. The methods in the first group are based on the detection of local intensity extrema as the top of the crowns; methods in the second group are based on the contours between the crowns and background; while the third group matches templates and looks for the best local correlations. Finally, we present new, interesting works published very recently. Exhaustive overviews can be found in Brandtberg
Figure 4.4: Three typical cases of segmentation of individual tree crowns.
(1999), Erikson (2004), Perrin (2006), and Shao and Reynolds (2006).
The simplest group of tree crown delineation models uses local intensity maxima to identify the most illuminated, i.e. the brightest, part of the crowns, which is usually the tree top. Local maxima are pixels which have greater intensity than all the other pixels in a defined neighbourhood. Blazquez (1989) introduced a method for finding citrus trees using local maxima to identify tree crown centres, combined with crown perimeter, mean intensity and area. He also presented a method for separating joined trees, using shadows and radiometric normalization between photographs. The aerial images are often affected by noise, resulting in bright pixels and misleading the local maximum search algorithms.
Notice that the difference between this situation and tree crowns, which is a bright circular area with usually 10–30 pixel diameter, is that in the crown case a bigger bright area is present. Therefore, one idea for handling this situation is to convolve the image with a Gaussian-kernel. Dralle and Rudemo (1996) presented a model to find the appropriate parameters. The model is based on smoothing the images with different filters,i.e.different scales, and comparing the result with field plot data of the number of stems present.
The second group of methods is based on contours. The term ‘contour’ in image pro-cessing is used as a delimiter between homogenous areas, but we note that in tree crown detection we distinguish contours between tree crowns and the background and valleys tak-ing place between individual tree crowns. Gougeon (1995) presented an approach to tree crown delineation. He separated the crowns from the background vegetation and from each other. The tree crowns are subsequently delineated using a five-level rule-based method designed to find circular shapes, but with some small variations permitted. Using MEIS-II images of coniferous plantations with a resolution of 31 cm/pixel, 81 % of the crowns are the same as those obtained by visual interpretation of the imagery. Brandtberg and Walter (1998) decompose an image into multiple scales (Lindeberg, 1998), and then define tree crown boundary candidates at each scale as zero crossings with convex greyscale curvature.
Edge segment centres of curvature are then used to construct a candidate tree crown region at each scale. These are then combined over different scales and a final tree crown region is grown.
Using templates to model tree crowns is a well studied and successful part of the field.
Although template matching methods are known as a classical searching technique, their success in the field is due to all the necessary information being available about aerial im-ages to create a very precise tree model. The first vision system capable of recognizing individual tree crowns, based on matching of a synthetic tree crown image model with an
aerial image, was developed by Pollock (1996). The system was tested on monocular high spatial resolution image data in Ontario and Alberta (Pollock, 1998). The procedure is based on a model of the image formation process at the scale of an individual tree. Natural vari-ation of the tree crown is considered, as is the species. Larsen (1998, 1999) concentrated on spruce tree detection using a template matching method. The main difference between these two papers is the use of multiple templates in the second. The 3D shape of the tree is modelled using a generalized ellipsoid, while illumination is modelled using the position of the sun and a clear-sky model. Reflectance is modelled using single reflections, with the branches and needles acting as scatterers, while the ground is treated as a Lambertian sur-face. Template matching is used to calculate a correlation measure between the tree image predicted by the model and the image data. The local maxima of this measure are treated as tree candidates, and various strategies are then used to eliminate false positives.
More recently, new approaches have been proposed using novel ideas and mixing the three previously used image data. Erikson (2003) used a region-growing method to sepa-rate crowns of mixed forest in high-resolution aerial images in central Sweden. The method starts from single (seed) pixels as representative pixels on the crown and spreads over neigh-bouring pixels that satisfy spatial and spectral requirements. In Erikson (2003), the region-growing part was completed with a random walk. The methods described so far use multiple steps rather than a unified model. Closer in spirit to the present work is that of Perrin et al.
(2004, 2005), who model the collection of tree crowns by a marked point process, where the marks are circles or ellipses. An energy is defined that penalizes, for example, overlap-ping shapes, and controls the parameters of the individual shapes. Reversible Jump Markov chain Monte Carlo and simulated annealing are used to estimate the tree crown configu-ration. Compared to our model, the method has the advantage that overlapping trees can be represented as two separate objects, but the disadvantage that the tree crowns are not precisely delineated due to the small number of degrees of freedom for each mark.