Leaf area index (LAI) describes the main plant surface area for gas exchange. Accurate LAI measurements are integral to effective hydrological, ecological, and climate modelling. LAI is commonly modelled using canopy gap fraction measurements from optical sensors. In woody vegetation, however, the wood to total plant area ratio (alpha) must also be estimated to convert plant area index (PAI) to LAI. Historically, estimating alpha required destructive harvests and is a potential source of LAI error. In this study, we present a theoretical framework for estimating LAI from digital hemispheric canopy photography by correcting for alpha within each image using semi-supervised pixel classification. We apply this framework to 201 images collected in semi-arid Australian vegetation (overstorey LAI range 0-5) to explore potential sources of error from: image classification, LAI model implementation, and differences in alpha among vegetation types. Leaf, wood, and canopy gap (sky) pixels were classified using a random forest (RF) algorithm with 87.7 +/- 0.01 % accuracy (mean +/- standard error) under overcast skies but 81.3 +/- 0.01 % under clear sky conditions where leaf and wood pixel classification was inconsistent. LAI estimates using the proposed approach had a strong linear relationship to PAI (r(2) >= 0.97). However, the proportional contribution of woody material to canopy gap fraction was zenith angle dependent. Allowing alpha to vary by zenith and azimuth angle when calculating LAI resulted in estimates 10-17 % higher than widely used PAI conversion methods. The zenith angle distribution of alpha also differed among co-occurring vegetation types. Allowing the PAI to LAI regression slope to vary based on the dominant genus reduced PAI conversion error by similar to 2 % (p < 0.001). Quantifying alpha variability within canopies and between vegetation types using the method outlined here can reduce on-ground LAI measurement uncertainty.