Model-based local thresholding for canopy hemispherical photography

Abstract

Canopy hemispherical photography (HP) is widely used to estimate forest structural variables. To achieve good results with HP, a classification algorithm is needed to produce binary images to accurately estimate the gap fraction. Our aim was to develop a local thresholding method for binarizing carefully acquired hemispherical photographs. The method was implemented in the R package “caiman”. Working with photographs of artificial structures and using a linear model, our method turns the cumbersome problem of finding the optimal threshold value into a simpler one, which is estimating the digital number (DN) of the sky. Using hemispherical photographs of a deciduous forest, we compared our method with several standard and state-of-the-art binarization techniques. Our method was as accurate as the best-tested binarization techniques, regardless of the exposure, as long as it was between 0 and 2 stops over the open sky auto-exposure. Moreover, our method did not require knowing the exact relative exposure. Intending to balance accuracy and practicality, we mapped the sky DN using the values extracted from gaps. However, we discussed whether a more accurate but less practical way to map sky DN could provide, along with our method, a new benchmark.