Lidar360 Tree Segmentation Case Study in Port Moody, British Columbia

Sustainable forest management requires a robust forest inventory. Luckily the advancements in UAV lidar systems are allowing for accurate extraction of forest metrics. Traditionally, forest metrics (including tree height and diameter at breast height (DBH)) are measured in the field and are very labour intensive to extract, especially in complex terrain. Instead, UAV LiDAR is increasingly becoming more insightful and cost-effective.

Figure 1.1: UAV LiDAR Vegetation structure detailed in a cross section view

Forestry management is generally divided into either area-based management or individual tree management. Here in British Columbia, area-based management is generally the most practical due to the vastness and density of forested lands. But as technology advances, the level of detail we can extract from the LiDAR scan is opening up new doors for individual tree inventories. With this in mind, we set out to see what sort of individual tree segmentation results we could achieve, even in such a dense temperate rainforest.

The project site was approximately 0.13 km^2 (~32 acres). This Pacific west coast location consists mainly of Westen Hemlock, Grand Fir, Western White Pine and Red Alder. The survey used Green Valley International’s (GVI) S220 LiDAR system mounted to a DJI M600 UAV. The survey was collected in a single flight (~10 minutes) at 80 metres AGL at a speed of 4 m/s.

Figure 1.2: Colourised point cloud output

Methodology:

The project area consists of forest, grass, low-density residential and a few paved roads. The survey had an average density of 381 points/㎡. For accuracy assessment, 30 ground control points (GCPs) were collected across the area using a Trimble Catalyst GNSS receiver. To help assess GVI Lidar 360 tree segmentation, manual field measurements of tree height and diameter at breast height (DBH) were extracted for 12 different trees. Tree height was extracted using an angle calculator and a tape measure, while DBH was recorded using a tape measure. After post-processing of the laser, IMU and GNSS data was done, further processing in Lidar360 included a boresight strip alignment, cutting overlapping points between strips and removing outliers. Ground points were then classified. The point cloud was then tied into the ground control, using half of the surveyed hard surface GCP measurements. Then an accuracy assessment was conducted using Lidar360’s control point report tool using the other half of the hard surface GCP measurements. Then a digital terrain model (DTM), a digital surface model (DSM) and a canopy height model (CHM) were created for the CHM tree segmentation. In addition to the tree segmentation, various other tree metrics were extracted from the point cloud, including tree height at different percentiles, canopy cover and leaf area index (LAI)

Figure 1.3: Port Moody forest scan site processing Individual tree segmentation in Lidar360.

Results:

The vertical root mean squared error (RMSE) was 0.051 cm, meaning that the point cloud has 5cm absolute accuracy to the real-world position of the LiDAR measurements. Individual tree identification assigned a unique number to every tree. The tree segmentation results show that the segmentation method tends to over segment. Using a 1m spatial resolution DEM, DSM and CHM (instead of 0.5m) helped minimize over-segmentation. The derived tree heights from the LiDAR data was significantly correlated to the tree height measurements collected in the field, resulting in a maximum error of 1.8 m, a minimum error of 10 cm and Average error of 17 cm.

Discussion:

The tree segmentation performed very well for coniferous tree species, even in areas where coniferous canopies overlapped. However, tree segmentation was much less robust for deciduous trees. This was partly due to the data being collected during winter, in leaf-off conditions. This resulted in the tree segmentation interpreting many of the bare branches as individual trees. Overall we found that tree height for conifers can be efficiently derived from UAV-LiDAR data. While the point density (~381 points/㎡) of the scan gave good detail of the majority of the forest's structure, there is more potential for better DBH value extraction from UAV-LiDAR with higher point densities.