COS 268-1 - Automated point cloud segmentation and novel connectivity metrics for vegetation classification across diverse forested ecosystems

Abstract: The assessment of vegetation connectivity is of great importance for monitoring habitat, assessing wildland fuels and understanding ecohydrology, especially under a changing climate. Techniques traditionally employed to measure vegetation structure include both indirect, rapid visual assessments and detailed transect approaches. With advances in technology, these on-ground approaches have recently been supplemented by remote sensing approaches utilizing LiDAR and 3D photogrammetry capture techniques to produce high-resolution 3D data (point clouds) and new metrics to describe connectivity (e.g. point density between certain heights). Prior research has largely focused on replicating existing metrics and do not consider the different vegetation components nor available information content within the point cloud. The goal of this presentation is to address this challenge by applying a deep learning approach (Forest Structural Complexity Tool) to separate the point cloud into vegetation, stems, coarse woody debris (CWD) and terrain in 5 distinct forest types: Mallee desert (Australia), Wet Eucalypt forest (South-eastern Australia), a conifer-dominated forest (Pacific North West USA), Sequoia forest (California, USA) and Lodgepole Pine landscape (South-Eastern USA). Subsequently, we also present two novel connectivity metrics (average distance between vegetation clumps and volume of vegetation clumps) that utilize the segmented point clouds in combination with a novel 3D alpha shape algorithm.

The results from our study highlight a high level of classification accuracy in comparison to manually separated point clouds across all forest types (classification accuracy: >70%). Eucalypt forests were classified most accurately (classification accuracy: 87%) due to the historical training datasets used in the development of the FSCT model. Across all sites, the distance between vegetation clumps was seen to have the highest correlation to existing connectivity metrics of Canopy Base Height and Fuel Strata Gap. Whilst strong correlation is seen to existing metrics, we argue that our novel 3D connectivity metric is more broadly applicable across forest types in comparison to point densities between certain heights which require tuning. Furthermore, the 3D connectivity metrics developed in this study provide a greater understanding of how vegetation is distributed both horizontally and vertically in comparison to existing metrics of FSG and CBH which only consider the vertical arrangement of vegetation. Overall, the utilization of metrics which leverage the full information content within point clouds will provide a novel pathway to describe vegetation structural characteristics through time and provide valuable information for ecological applications such as ecohydrology, habitat assessments and vegetation condition pre and post fire.