Developing An Enhanced Forest Inventory In Maine Using Airborne Laser Scanning PDF Download

For more than a century, forest inventories have been used to support forest management and timber valuation activities. Today's inventories still rely primarily on manual measurements combined with sampling and modeling techniques. In recent years, new opportunities in carbon sequestration and an increasingly sophisticated timber market have prompted a need for more scalable and efficient inventory methods. To meet this demand, the industry has turned to remote sensing--predominantly light detection and ranging (LiDAR), which utilizes lasers to scan and measure features in 3D space. While much progress has been made, data resolution and cost challenges for both airborne and terrestrial LiDAR still exist. Airborne Laser Scanning (ALS) is more efficient for measuring large forest areas but faces challenges with respect to resolution and occlusion, leading to omission of understory trees. Terrestrial Laser Scanning (TLS) performs better in those respects but relies on expensive and typically unwieldy hardware. Area-based LiDAR approaches have been successful for large-scale applications but are not ideal for smaller parcels. As such, a need persists for a LiDAR-based solution that enables efficient generation of large-area forest inventory data yet is scalable to smaller forest plots and a range of forest types. The objective of this study was the development of a scalable individual tree detection method that leveraged airborne LiDAR data and performed well in mixed-species hardwood forests found in the northeastern United States. Existing research in individual tree detection has focused on methods that work well in conifer-dominated forests and homogenous settings such as plantations. These approaches, many of them based on top-down canopy height models, perform less favorably in deciduous stands due to the canopy complexity and crown characteristics inherent to these forest types. The voxel-based method proposed here uses detailed ground-measured tree survey data and leaf-off LiDAR collected in 2019-2020 over the Shavers Creek Watershed in Pennsylvania, United States. The method detected 68% of all reference trees greater than 10cm diameter at breast height (DBH) and 87% of sawtimber-sized trees greater than 28cm DBH, and it performed consistently across 48 subplots in the three-hectare test area. A new tree matching method leveraging linear integer programming was used for training and evaluation of the method. This tool enabled true one-to-one matching of predicted and reference trees and the validation of tree detections. Mean positional accuracy for predicted trees was within one meter of ground-measured reference trees. The results indicate the method has potential to be operationalized for both traditional forest management activities and in meeting the demand for more frequent and scalable inventories spurred by a growing forest carbon sequestration industry.