"Quantifying leaf area index from plant area index in Australia’s dryland vegetation using semi-supervised classification of digital hemispheric canopy photographs"Eckersley, JakeLeaf area index (LAI) describes the main vegetative surface for gas exchange which strongly regulates land-surface atmosphere interactions. Accurate and timely LAI estimates are therefore integral to effective hydrological, ecological, and climate modelling. Canopy gap fraction (Pgap) models are applied extensively to measure LAI using both active and passive optical sensors and form the basis of satellite LAI product validation. However, quantifying the impact of woody elements on indirect Pgap measurements in forests and woodlands has previously relied on laborious and destructive manual sampling and is a significant source of measurement error. In this study we present a novel framework for separately measuring LAI and plant area index (PAI) from digital hemispheric canopy photographs (DHCP) using a semi-supervised image classification workflow. We then applied this framework to 200 DHCP images collected in vegetation groups typical of semi-arid Australia to a) explore predictors of classification accuracy and b) quantify the relationship between LAI and PAI. Utilising a random forest algorithm, we separated leaf, wood, and canopy gap pixels with a mean accuracy of 93.6% (± 0.03%) under clear sky conditions. Accuracy decreased accuracy under overcast skies (83 ± 0.06%; p = 0.001) due to poorly differentiated leaf and wood elements. Woody material accounted for 45% of plant surface area measured by DHCP. Whilst previous studies have used this ratio to convert PAI to LAI, we found this approach underestimated LAI by 42% when leaf occlusion was estimated. PAI was a strong predictor of LAI (r2 = 0.97; p = 0.001), however, we found significant variability in the relationship slope between major vegetation groups typical of dryland Australia (p = 0.001) and accounting for this variability reduced prediction error by 27% (mean = 0.07 LAI units). The non-destructive approach outlined here provides an alternative for estimating LAI from DHCP, prompting future research exploring alternative image classification models to reduce uncertainty. Additionally, robust canopy simulations quantifying leaf and wood element occlusion may further refine LAI modelling assumptions. |
« back