An efficient approach to standardizing the processing of hemispherical images for the estimation of forest structural attributes

Digital hemispherical photography (DHP) has become a widely used tool for the estimation of forest structural attributes, such as gap fraction, Leaf Area Index (LAI), effective Plant Area Index (PAI(e)), and clumping. This development was boosted not only by a rapid technical advance in the field of digital photography but also by the inherent advantages of DHP for in situ measurements of forest structural attributes. However the major drawback of using DHP for the estimation of forest structural attributes is the lack of standardization which impedes a consistent compatibility with other indirect methods. This lack of standardization is mainly due to uncertainties introduced at the stage of image acquisition and processing. Of these, the determination of optimum exposure and thresholding in the image processing chain are two major influences. In this work influences on the estimation of forest structural attributes, namely the radiometric image resolution, the file format and the image band selection, were studied, in particular with regard to the inter-dependence with exposure and the threshold algorithm applied. For this purpose four different automatic threshold algorithms (Ridler, Otsu, Minimum, Isodata) were tested. Results show that the file format and the image band selection influence the estimation of gap fraction, PAI(e) and clumping indices. The magnitude of this effect however varies with the threshold algorithm applied, i.e. with a strong effect for the Minimum and Isodata algorithms and little effect for the Ridler and Otsu algorithms. The radiometric image resolution was found to cause only a marginal effect. Based on a comparison with LAI-2000 measurements it could also be demonstrated that the file format and the image band selection affect the determination of the optimum exposure. To resolve these issues an efficient approach to standardizing the processing of hemispherical images is proposed. This approach constitutes the stacking of five differently exposed hemispherical images and passing them to an automated clustering algorithm (Isodata) with the subsequent generation of gap fraction images. The resulting PAI(e) estimation performs better than or comparably to the estimation based on optimally exposed single images. In addition to being robust and objective, our approach provides consistent compatibility with the LAI-2000. (C) 2012 Elsevier B.V. All rights reserved.