William Woodgate

Derivation of Leaf Area Index and Associated Metrics from Remotely Sensed and In Situ Data Sources

Will Woodgate Squared
University
University of Melbourne
Supervisor (Academic)
Prof Simon Jones, RMIT & Prof Joe Leach, University of Melbourne
Supervisor (Industry)
Andrew Haywood, DEPI Vic
Projects
mysite
Employment
Research Fellow at CSIRO in the Oceans and Atmosphere Flagship
Thesis Abstract

Leaf Area Index (LAI) is an essential climate variable functionally related to the energy and mass exchange of water, carbon, and light fluxes through plant canopies. It is defined as half of the total leaf area per unit ground area. LAI is commonly derived from a number of active and passive remote sensing instruments on satellites, aircraft and on the ground. There is an increasing need for more accurate and traceable measurements in support of calibration and validation of Earth Observation (EO) products. Ambitious accuracy targets as low as 5% error are specified by the Global Climate Observing System (GCOS) and associated end-users. This poses a challenge for commonly used remote (indirect) retrieval techniques, which typically suffer from a greater level of uncertainty than direct methods. On the other hand, indirect methods are preferred over direct methods due to their scalability and cost effectiveness compared with manually-intensive, costly and destructive methods for the attribution of plant communities.

This research set out to examine means to improve uncertainty in the estimation of LAI in forests. It specifically sought to quantify uncertainty associated with indirect estimation of LAI from the application of the ubiquitous Pgap physical model (Monsi & Saeki, 1965; Nilson, 1971). The physical model calculates LAI from physically quantifiable factors of gap probability (Pgap), canopy element clumping, canopy element (leaf and wood) angle distribution, and the proportion of wood-to-total plant area ‘α’. All of these metrics are required to be estimated or assumed to within an acceptable margin of error for LAI estimation.

This thesis was conducted in three stages. Stage 1 compared data collection and processing methods following standard operational procedures in five diverse forest systems yielding LAI values ranging from 0.5 to 5.5. Data were collected synchronously and coincidentally from a Riegl VZ400 terrestrial laser scanner (TLS), high- and low-resolution digital hemispherical photography (DHP), and an LAI-2200 plant canopy analyser. A high degree of variance was found between these systems and subsequent processing methodologies; more than half of the pairwise comparisons had an RMSD ≥ 0.5 LAI, and one third were significantly different (p < 0.05). These results demonstrate that the variability between commonly utilised indirect ground-based methods need to be further reduced in order to provide repeatable unbiased and accurate validation estimates to meet product accuracy targets as low as 5%. Recommendations and guidelines for data collection and processing were developed, in addition to suggestions that could lead to reduced variability via TLS calibration and improved DHP image capture and processing methods.

However, the main impediment for assessing LAI method accuracy was the lack of a precise benchmark or true value, which is unattainable in a forest environment. Therefore in stage 2, a 3D modelling framework was developed to address this fundamental limitation. This framework was parameterised using a 3D scattering model coupled with 3D explicitly reconstructed tree models representative of a sampled forest stand, the first of its kind for an Australian forest. The 3D modelling framework enabled validation of the woody element projection function ‘Gw’, a newly proposed parameter in this study required to increase LAI accuracy through the application of the Pgap physical model. Gw characterises the angular contribution of non-leaf facets in woody ecosystems. Subsequently, a modification of the physical formulation is presented to include Gw, which directly links to an updated formulation of the extinction coefficient. LAI errors up to 25 percent at zenith were found when ignoring Gw and were shown to be a function of view zenith angle. The inclusion of Gw was found to eliminate this error.

LAI estimation sensitivity of the 3D models to leaf angle distribution (LAD) and its impact on within-crown clumping were investigated for the first time during stage 2. LAD was shown to considerably affect within-crown clumping levels of reconstructed tree models at nadir. However, at the 1 radian view zenith angle, within-crown clumping for individual tree models was largely independent of LAD. Within-crown clumping factors for the modelled dataset were as low as 0.35. Consequently, making a common assumption of a random distribution of canopy elements would lead to an LAI error of up to 65% for the modelled stand.

At stage 3, the 3D modelling framework was then extended to the simulation of DHP at the forest stand level, utilising a range of structurally diverse virtual scenes varying in stem distribution and LAI. This enabled validation of angular clumping retrieval methods, based on gap size distribution and logarithmic averaging approaches. The combined Chen & Cihlar (1995) and Lang & Xiang (1986) method from Leblanc (2002) was the best performing clumping method. It matched closely with the model reference values at nadir, with a linearly increasing error of greater than 30 percent PAI at the 75° view zenith angle. The framework was also applied to benchmark for the first time an indirect method to estimate the woody correction factor ‘α’ to convert plant area index (PAI) to LAI. The indirect ‘α’ method utilising classified DHP imagery matched to within 0.01 α of the reference, thus demonstrating its applicability for accurate indirect estimation in evergreen forests. The errors obtained when ignoring the effects of clumping and α in the representative virtual forest stand were as high as 55% and 45% LAI, respectively. On the other hand, the error was reduced to 6% LAI when applying the best performing clumping method and α retrieval method.

The findings of this study and the extended physical formulation presented here-in are applicable to sensors of all platforms calculating LAI from the Pgap physical model. They are especially relevant to clumped canopy environments or canopies where woody (non-leaf) elements contribute to the extinction of light.