Assessment of forest canopy vertical structure with multi-scale remote sensing: from the plot to the large area

Assessment of vegetation over large, remote and inaccessible areas is an ongoing challenge for land managers in Australia and around the world. This research aimed to develop metrics, techniques and acquisition specifications that are suitable for characterising vegetation across large forested areas. New methods were therefore required to be transferable between forest types as well as robust where forest structure is unknown a priori. Remote sensing techniques were utilised as they have been previously identified as key in forest assessment, owing to their synoptic capture as well as relative cost. Additionally, active remote sensing instruments, such as LiDAR, are capable of sensing 3-dimensional canopy structure.

Canopy height and the canopy height profile are fundamental descriptors of forest structure and can be used for estimating biomass, habitat suitability and fire susceptibility. To investigate the ability of remote sensing to characterise vegetation structure across large areas, three key research questions were formulated:

I. Which metrics of canopy height and vertical canopy structure are suitable for application across forested landscapes?

II. What is the appropriate ALS sampling frequency for attribution of forest structure across different forest types?

III. How can plot level estimates of canopy structure be scaled to generate continuous large area maps?

A number of inventory measured canopy height metrics were compared with LiDAR analogues, these were shown to be accurate at estimating canopy height and transferable between forest types. Existing techniques for attributing the canopy height profile were found to be inadequate when applied across heterogeneous forests. Therefore a new technique was developed that utilised a nonparametric regression of LiDAR derived gap probability that identified major canopy features e.g. dominant canopy strata and shade tolerant layers beneath.

The impact of sampling frequency was assessed using three key descriptors of canopy structure at six sites across Australia covering a range of forest types. The research concluded that forest structure can be adequately characterised with a pulse density of 0.5 pulses m-2 when compared to a higher density acquisition - 10 pulses m-2. At pulse density of <0.5 pulses m-2, the inability to generate an adequate ground surface model lead to poor results, particularly in high biomass forest. The outcomes of this research will allow land managers to specify lower pulse densities when commissioning LiDAR capture, which may result in significant cost savings.

Finally, LiDAR derived plot estimates were scaled to an area of 2.9 million hectares of forest, where forest type ranged from short, open woodland to tall, closed canopy rainforest. Attribution was achieved using a two-stage sampling approach utilising the ensemble regression technique Random Forest. Predictor variables included freely available datasets such as Landsat TM and MODIS satellite imagery. Canopy height was estimated with a RMSE of 30% or ~5.5 m when validated with an independent forest inventory dataset. Attribution of the canopy height profile was less successful for a number of reasons, for example, the relatively high spatial variability of shade tolerant vegetation. Inclusion of additional synoptic datasets, such as radar, may improve this in the future.