Automatic thresholding for hemispherical canopy-photographs based on edge detection

https://doi.org/10.1016/j.agrformet.2004.10.002Get rights and content

Abstract

The analysis of hemispherical photographs is nowadays an established method for assessing light indirectly and describing canopy structures. In this article, we present an automatic threshold algorithm for separating canopy and sky by edge detection. The algorithm was evaluated under different canopy conditions by comparing its results for canopy openness, fractal dimension and diffuse transmittance with those from multiple manual thresholding and direct measurements of the percent photosynthetic photon flux density (PPFD). We show that the automatic threshold algorithm is appropriate to replace the widely used manual interactive processing. It also improves the accuracy of results, especially in comparison with single manual thresholding. Whereas manual threshold setting has often been criticised as subjective and a major source of error the less time-consuming edge detection approach is objective, reproducible and can be applied to a large number of images.

Introduction

Light availability is an important site factor in ecology and the management of forests or agricultural land. It can be assessed indirectly in hemispherical photographs, a technique first used for forests by Evans and Coombe (1959) and Anderson (1964), which is nowadays widely applied (recently Bellow and Nair, 2003, Courbaud et al., 2003, Dignan and Bren, 2003, Halverson et al., 2003). With technological progress, several approaches for automation, especially computerised methods, have been proposed (Bonhomme and Chartier, 1972, Olsson et al., 1982, Chan et al., 1986, Chazdon and Field, 1987, Becker et al., 1989). In addition, several software packages for the image analysis are now available (Frazer et al., 2000), and the change from black and white film to digital camera systems has been evaluated (Englund et al., 2000, Frazer et al., 2001, Hale and Edwards, 2002).

Although using hemispherical photographs has several advantages over direct light assessment and technological and methodological progress have led to improvements in the method, it also has limitations. One of the most critical points is separating the canopy from the sky during image analysis. The usual initial step is a binary transformation generally done by setting a threshold manually, although there are some approaches that work directly with grey or colour values (Olsson et al., 1982, Wagner, 1998, Wagner, 2001). Several authors have pointed out that manual thresholding can be a relevant source of error because it is somewhat arbitrary and subjective (Chan et al., 1986, Rich, 1990, Machado and Reich, 1999, Frazer et al., 2001, Diaci and Thormann, 2002, Jonckheere et al., 2004).

In this article, we present an automatic threshold algorithm for hemispherical canopy-photographs based on edge detection. The algorithm is evaluated under different canopy conditions by comparing it with single and multiple manual thresholding and direct radiation measurements. The results are discussed in relation to manual thresholding and the known limitations of using hemispherical canopy-photographs to assess radiation indirectly.

Section snippets

Optimal threshold algorithm

The principle of the method is the search of a threshold value that gives highest local contrast at the edges between classified canopy and sky. Accordingly, the threshold value t with the maximum mean brightness difference at the edges is defined as optimal threshold topt (Eq. (1); Fig. 1).topt=argmaxt(meanS*{|bx1,y1bx2,y2|f(bx1,y1,t)f(bx2,y2,t)})f(b,t)=1:b>t0:btThe image domain S* of calculating the mean is defined as:S*={(x,y)|x[1,xmax1]y[1,ymax1]},with(x1,y1,x2,y2){(x,y,x+1,y),(x,y

Results

Fig. 2 shows the comparison of single manual and automatic thresholds with the mean manual threshold values. There are striking scatters for both methods showing similar magnitude of variation. However, the automatic algorithm tends to over- or underestimate the thresholds for the majority of images, this is particularly noticeable for the conifer forest and the uncleared windthrow subsets.

Results for the canopy openness and fractal dimension are given according to the different thresholding

Discussion

As shown in Fig. 2, the single manual thresholds vary greatly, which means that they are a potential source of error. Even more parameters calculated after single manual thresholding may vary remarkably too (Fig. 3). The influence of the variation in single manual thresholds on subsequent analyses depends on both the parameter and the image properties. After transformation, photographs with a high contrast between the vegetation and sky will show only slight differences in a wide range of

Conclusions

The results show that the suggested automatic threshold algorithm by edge detection has advantages over manual interactive processing. It is objective, comprehensible and reproducible, whereas manual thresholding has often been criticised as subjective and a major source of error. The edge detection approach may also improve the accuracy of the results. But its main advantage is that it is much less time-consuming than manual thresholding and can be applied to a large number of images. For

Program source

The automatic threshold algorithm is implemented in the software tool SideLook. It can be downloaded as shareware at http://www.appleco.ch.

Acknowledgements

We are especially grateful to J. B. Stewart and two anonymous reviewers for their valuable comments on the manuscript. The authors would like to thank S. Martin for providing the conifer forest images. We are grateful to C. Dähler, L. von Fellenberg, A. Ghiringhelli, R. Häner, M. de Montmollin and M. Zubler for their help with manual thresholding. We also thank C.W. Hoffmann for his help with mathematical equations and wording. This work was supported by the Swiss Agency for the Environment,

References (31)

  • S. Wagner

    Relative radiance measurements and zenith angle dependent segmentation in hemispherical photography

    Agric. For. Meteorol.

    (2001)
  • A. Zehm et al.

    Multiparameter analysis of vertical vegetation structure based on digital image processing

    Flora

    (2003)
  • M.C. Anderson

    Studies of the woodland light climate. 1. The photographic computation of light conditions

    J. Ecol.

    (1964)
  • R. Bonhomme et al.

    Interpretation and automatic measurement of hemispherical photographs to obtain sunlit foliage area and gap frequency

    Isr. J. Agric. Res.

    (1972)
  • S.S. Chan et al.

    Evaluating forest vegetative cover with computerized analysis of fisheye photographs

    Forest Sci.

    (1986)
  • Cited by (225)

    View all citing articles on Scopus
    View full text