Mapping understory vegetation in the Boreal forest: Seeing challenges as learning opportunities.
The overarching goal of my master’s project is to map and characterize the understory vegetation in the boreal forests of northern Alberta. The understory is an interesting topic of study because it has been largely disregarded as noise in many forestry studies using remote sensing, since it does not hold the same economic value that timber producing trees do. However, this vegetation layer is an important component of the forest because it influences ecosystem processes such as fire regimes, nutrient cycling, and wildlife habitat (Figure 1). Increasing our understanding and quantifying the understory layer can give us more insight into the mechanisms behind those ecosystem processes.
Figure 1. Ecosystem processes influenced by understory vegetation.
The way that I am characterizing the understory vegetation is by trying to understand the structural arrangement of plant material in space. My approach is to look at the understory vegetation from a different angle. This angle observes plants as objects that are made up of matter and occupy space. Objects also have physical characteristics that help describe their physical structure. Therefore, I am trying to explain this by breaking down vegetation structure into simplified physical attributes. There was a total of 5 physical attributes or metrics that I am using to characterize the understory, these include height, % cover, density, complexity, and volume (Figure 2).
Figure 2. Physical attributes used to understand understory vegetation structure.
Remote sensing technology such as LiDAR is ideal to use in this study because it collects 3-dimensional structural data of the forest plant material (Figure 3). It also covers a large spatial extent, which enables the application of understory characterization methods over the substantial areas covered by LiDAR.
Figure 3. (Left) Depiction of aerial LiDAR sensor emitting laser pulses. (Right) 3D LiDAR point cloud of vegetation found in a field plot.
Since one set of data is in the format of a point cloud, we wanted to structure the field data to have the same 3-dimensional point format. This is useful because it allows me to apply the same metric extraction approach to both the field and LIDAR point clouds in the same way. As a result, a field methodology was designed to be able to emulate the way that LiDAR collects information by setting up field plots that were subdivided into a grid of points 1-meter spaced apart. At each grid point we recorded the height and species of plant objects (Figure 4).
Figure 4. (Left) Depiction of sampling in the field. (Right) Example of height measurement pole used to measure the height of vegetation at grid intersections.
The LiDAR data was able to capture the tall canopy and understory alike, so we had to limit the height threshold for this study because it would be difficult to sample plants in the field that were really tall. As a result, the understory vegetation in my study is defined as any vegetation that has a height range between 10 cm and 3 meters. The lower range value was selected because in the LiDAR point cloud, those points would be difficult to distinguish from the ground points. The 3-meter threshold was selected because the browsing height of many herbivorous species, such as deer and caribou, ranges from 1.5 – 2 meters.
The output of the field and remote sensing data collection results in two sets of point clouds for each of the field plots (Figure 5). The physical attributes were derived at the plot-level instead of at the point-level, because many of the metrics require all of the points to be included in their calculation.
Figure 5. Example of field and LiDAR point clouds that were used to derive understory attributes, they were both limited to a 3 meter height range.
The preliminary results of this study indicate that there is an overall bias in the relationship between the field and LiDAR derived metrics (Figure 6). More specifically, the plot-level summaries based on the LiDAR point cloud underestimate summary values from plots calculated from the field data. This behaviour was observed for most physical attributes we derived. In addition, patterns associated with canopy closure and ecosystem type can be observed. Plots that are underestimated the most tend to have high to considerable canopy closure (i.e., think dense forest), and are located in the uplands. On the other hand, plots that are located in the wetlands and have low to moderate canopy closure are underestimated to a lesser degree.
Figure 6. Scatterplots showing the results found in the comparison between field and LiDAR point clouds.
I decided to investigate further. An idea occurred to me which relates back to how the two point clouds were collected. An interesting difference between the two point cloud sampling approaches was noted. The LiDAR samples the vegetation from above and the field samples data from the ground, and this influences the height of the points that are present in either point cloud (Figure 7). The points that are unoccluded (i.e., not covered by tree canopy) get sampled in both the LiDAR and the field, whereas a point that is heavily occluded by a canopy, will not be sampled in the LiDAR. In general the occluded points that were not sampled by the LiDAR include plant objects such as tree trunks, which can reach the 3 m height threshold. This means that the field-based plot summaries include more points at this height than the LiDAR-based plot summaries. This would potentially explain why the metric values derived from the field are higher than those in LiDAR.
Figure 7. Comparison of how occluded and occluded points are sampled by the LiDAR and in the field.
This was then tested by removing the points in the field point cloud that were tagged as tree species, including black spruce, jack pine, and aspen poplar. The new scatterplots no longer show a bias, but they also no longer show patterns associated with canopy closure and ecosystem type (Figure 8). This means that the occluded points that tend to have greater heights represent points recorded from trees that are present within the 3-meter height threshold. Trees trunks, for example, tend to have much taller heights that the surrounding vegetation, and are commonly occluded by the canopy above. The presence of these trees in the field data led to higher values that were being observed. However, the information from those trees is important to keep, because they still contribute to the physical structure of the understory layer. In addition, there is a connection between the trees and the other factors, being canopy closure and ecosystem type. Since trees lead to higher canopy closure, which was also shown to occur in the uplands.
Figure 8. Scatterplots showing the effect of tree point removal from the field data. The distribution of the points do not indicate a bias. The patterns related to canopy closure and ecosystem type are also no longer seen.
To conclude, we were able to explain the biases initially observed in the data, which were due to the occlusion of vegetation. Occluded tree material being sampled more often in the field data than by the LiDAR, led to the underestimation of field values by the LiDAR data. Patterns associated with canopy closure and ecosystem type showed that plots located in the uplands had greater degrees of canopy closure, and were also the ones that were most underestimated.
The next steps for my study include deriving the same metrics using a different approach that involves voxels, and determine if there is an improvement in the relationship with the field-derived understory attributes. Then the approaches with the strongest correlations will be selected for further analysis and map generation.
This cover has been designed using resources from Freepik.com. Tree graphics designed by macrovector.