UncertainTree

To estimate timber production, biomass, and other important values for foresters on large areas, laser scanning can be used. Typically, a laser scanner is mounted in an airplane or on a drone. An invisible laser beam is sent out from the instrument and is reflected on leaves, branches, stems, and the forest ground. These reflections are measured by the instrument and converted to a three-dimensional image of the forest, called a point cloud. Since the laser beam is not a perfect line, but diverges to illuminate an area, the leaves do not fully shield the forest floor and parts of the beam are reflected underneath the canopy. This allows to gather information on the vertical architecture of the forest. As trees move in the wind and change their leaves positions throughout the day and over seasons, the point cloud will look different every time it is recorded, even if the trees have not grown. This research projects wants to find out how much change in the calculated values for example in biomass can be expected. Knowing this uncertainty allows researchers and foresters to tell whether an increase in biomass, when measured from the point cloud, is just an effect of wind moving the leaves in a specific way, or if the trees have really grown since the last recording. To investigate these uncertainties, we will try three different approaches: First, we will look at parts of the forest that are scanned twice, with only a few minutes between the recordings. As laser scanning from airplanes is typically done in stripes, and these stripes overlap a bit to ensure that nothing is missed, we can look at these overlapping areas. Then, we quantify the biomass from both acquisitions separately (usually one would just merge the datasets) and see how much they differ. We then analyze these differences statistically. In the second experiment, we will look at forests that have been recorded multiple times over the last years. We try to teach a computer how trees grow by showing it many trees of the same species. We assume that they all grow similarly, if they are similarly old and l ive close to each other. Then we subtract this growth from the time series, and we should be left with exactly the uncertainty that is a result of the leaves being different than the last time. Finally, we create a 3D computer model of a couple of trees. Using a physical model, we apply wind force to the trees and take snapshots. Then, we simulate laser scanning acquisitions on these snapshots. This will allow us to see how much the different leaf and branch positions influence the calculated biomass. Since we can decide how strong the wind should blow, we can see how the wind influences the uncertainty in the quantification. We assume that a stronger wind will lead to higher differences. Furthermore, different tree species are assumed to have different resilience e.g., an aspen tree will move much more in the wind than a maple tree.

Biomass is a crucial parameter in assessing our forests. To obtain large-scale estimates of biomass, data from 3D point clouds obtained through airborne laser scanning are combined with ground-based point measurements from forest inventories. When investigating changes such as forest growth or stagnation e.g., due to pest infestations, it is essential that biomass values come with an estimation of their associated uncertainty. Only then can meaningful comparisons be made. In the FWF project "UncertainTree," researchers explored how to determine this uncertainty from existing data. Data from the Petawawa Research Forest in Canada were divided into individual flight strips. These strips were flown in an overlapping fashion, allowing for multiple quantifications of biomass at the same location. The strip-wise biomass values were then compared to results obtained without prior splitting. A root-mean-square-difference (RMSD), a measure of variance, was determined to be between 10-20 tons per hectare. In comparison, the model error in biomass estimation ranges from 40-50 tons per hectare (for total biomass values between 0-150 tons per hectare). This highlights the importance of considering the effects of data acquisition geometry and vegetation movement during data collection. Flight conditions should hence be reported in metadata. In a second experiment, an artificial neural network (NN) was used to directly derive biomass from the laser scanning point clouds. This was carried out for two epochs, specifically for the years 2012 and 2018 in the case of the Petawawa Research Forest. The neural network autonomously learned a feature vector suitable for biomass estimation, which was extracted and used to train another neural network, aiming to derive the vector obtained from the 2018 point cloud from the 2012 one. This corresponded to learning of a model of forest growth. Finally, difference between the biomass predicted for 2018 from the 2012 data and the actual biomass in 2018 could be determined. Here, the RMSD of 30-40 tons per hectare again provided a measure of how uncertain individual biomass estimates are. To separate the influence of data acquisition geometry from object movement during data collection, additional simulations were conducted. Highly detailed computer-generated 3D tree models were created, and the movement of the trees under different wind strengths was simulated. Snapshots were then used in a laser scanning simulation software to create point clouds. The average height of laser points, previously identified as an important parameter for biomass estimation in the first experiment, was determined for different wind strengths (with variations of up to 70 cm) and for different data acquisition geometries (with variations of up to 5 cm), and compared to the reference without wind. The simulation again demonstrated that the movement of vegetation during data collection can have a significant impact on biomass estimates.