2. OBJETIVOS
5.3 CONVIVENCIA EN EL AULA
5.3.2 Aprendizaje cooperativo
The first stage in the individual crown delineation process was the development of crown templates using field data derived empirical functions. The use of minima contouring and template matching methods to assist in crown delineation using high-spatial resolution imagery was described in Chapter 2 (e.g., Koukoulas and Blackburn, 2005; Gougeon and Leckie, 2006; Bunting and Lucas, 2006; Solberg et al., 2006), and crown shape was shown to influence LiDAR response (Nelson, 1997).
An examination of the Injune field data identified two broad categories of crown shape/area (i.e., a template) for taller mature trees – small compact crowns regardless of top- height (i.e., generally less than 50 m2), such as those from Callitris or Acacia genus (or some
Eucalyptus in heavily stocked stands). The other category had large wide spreading crowns as the maximum height of the tree was attained, which is typical of woodland Eucalyptus or Angophora genus. Shorter, less mature trees of both types generally had smaller, more compact crowns, with the mean crown area of all trees being less than 10 m2 (Figure 23).
Figure 22: Flowchart of crown delineation methodology Examine field data to
see which broad tree structure types are
present
Derive height to crown area relationships for the
two main tree structure types
Tall tree, wide crown function (Eucalypt) Tall tree, small crown function (Callitris)
Create tree / non-tree boundary from HSCOI contour that encompasses
~90% of LiDAR returns (2m+)
Create segmentation using HSCOI layers smoothed at 1m
and 5m scales. Smooth boundaries and remove slivers
Intersect tree crown boundary with crown segments from both 1m & 5m scale layers, and remove internal boundaries from non-crown
areas. Smooth boundaries, remove slivers, and merge very small segments
For each 1m & 5m scale crown segments, calculate tree structural type using LiDAR
vertical profiles
Union 1m and 5m scale crowns. Update 1m segment structural type by comparing to associated 5m segment type
Check shape of new segment is not too irregular (i.e. is broadly elliptical) by using a vector based ray trace in 8 directions, then apply four criteria that assess
ratios of radius length & area
Shape ok Keep removing segments until new area ≤
80% of starting area and;
- shape was ok, or until a single segment is left on list, or until checking loop has run 10 times
Shape not ok From the list of merged segments,
remove the segment that has centroid furthest away, and was least similar for height and HSCOI
Create final crown delineation using ’merge’ ID. Calculate final crown measurements, check with
field data Create new segment from
updated list of component segments
Apply four merge selection criteria to surrounding segments, using thresholds of
height, adjacency, distance between segment centroids and HSCOI value
Iteratively select surrounding segment centroids (tallest to shortest) within search radius of core segment centroid, as potential segments to merge with. Segments
need to have at least 10% area within search circle
Those segments which meet selection criteria are assigned the ‘merge’ ID of the core segment, and all are merged into one new crown
segment Those segments which don’t meet selection criteria are not allocated a
merge ID
Iteratively process each 1m scale crown segment (from tallest to shortest), if not already allocated this becomes core segment
for this iteration. Calculate expected crown area (+50%) using segment height with tree type function, then derive search radius from
this area (assuming circle)
Stop including segments when the sum area of current segments (core + any additional segments) was greater than the
From examination of the range of field data, a basic hypothesis was postulated that the taller the tree, the wider the crown should be, a proposition also put forward by Jack and Long, (1991). Different crown size ranges per tree height were observed for each of the two main crown shape/area categories (Figure 23). It was also observed that the crown area of these two different types could be significantly different for the same tree top-height, especially at the taller end of the observed range (i.e., > 20 m).
The crown delineation process utilises a multi-scale approach, with a fine scale segmentation to derive crown building blocks (equivalent to a Level –1 in a hierarchy), which were then compared to segmentations derived at medium scales, to put the constituent blocks into context (equivalent to a Level 0 in a hierarchy). The final crown products were a combination of constituent crown segments. The primary processing issue then becomes one of appropriately merging small segments into large (but not too large) segments for trees with wide crowns, whilst retaining small segments for smaller crowned trees. The appropriate sizes were determined from empirical functions derived from field data, which compare tree top height and crown area.
Crown area was based on an assumed circular or elliptical shape, derived from diameter measurements in north-south and east-west directions. The calibration and validation of both functions are shown in Figure 23, with calibration utilising a random selection of 80% of field stems, and validation using the remaining 20% of stems. The Eucalyptus function was described in Lee and Lucas (2007) and shown in Figure 23a, with the Callitris function given in Equation 3 (Figure 23b).
Callitris Crown Area = 0.6027*exp(0.2015* height) Equation 3
Using randomly selected validation datasets, the Angophora/Eucalypt function produced a correspondence (r2 = 0.70, RSE = 16.2 m2, n = 249), with a slope and intercept of the best-fit line being 1.66 and -1.937 respectively (Figure 23). The Callitris/Acacia function produced a correspondence (r2 = 0.38, RSE = 5.8 m2, n = 506), with a slope and intercept of the
the Callitris/Acacia function validation result, on average the function produced near to the expected field result (as evidenced by the validation best-fit line approximating the 1:1 line). Because the actual crown areas involved were relatively small (i.e. generally less than 50 m2), the function was considered adequate for the intended purpose within the delineation modelling.
a) b)
Figure 23: Calibration using 80% of field data trees (Upper) and validation using 20% of field data trees (Lower) for estimating crown area from height, for (a) Eucalypt and Angophora trees, and (b)
Callitris and Acacia trees (live trees 5cm+ D130).
However, the validation exercise indicated that the expected crown area derived using the Angophora/Eucalypt function was consistently underestimated when compared to the field estimated crown area. From the validation graph (Figure 23), the line -of-best-fit was found to show bias in the prediction of actual field measurement value of approximately 1.5 times the translation function value. Therefore to improve the predictive ability of the function when applied to the LiDAR modelling, a crown area expansion factor of 50 % was applied to the initial function result. In the modelling algorithm the final expected crown area calculated by
dividing the initial expected crown area by two, and the result added to the original expected area. The observed bias in the translation function validation was most likely a result of the large natural variation in Angophora and Eucalyptus height and crown area as measured in the field. This could be linked to different species, stem density influences, soils, nutrient, and water availability (Florence, 1996). Lack of accuracy in the crown area field measurements may also contribute to the observed variation in the relationship between crown area and height.
As the height-to-crown area function was exponential, very large crown areas can be generated at the upper end of recorded tree heights. To prevent unrealistic crown area values being produced, an upper crown area bound was applied to the final expected crown area. This upper bound was based on the largest crown area observed in the field data, plus an expansion factor of approximately 10 % to allow for error in the field estimate, and for trees larger than those found in the field plots. For the Queensland data, the largest crown area observed was 455 m2, so the upper bound applied was rounded up to 500 m2. With the Queensland derived
function, the maximum crown area would be derived with a tree height of approximately 27 m or more. When considering all LiDAR in the landscape sample, tree heights up to 35 m were observed, whilst the tallest tree measured in the field was 31 m, so it was therefore possible that larger crowned trees exist in the landscape. Examination of LiDAR (or other high spatial resolution data if available) at the locations with the tallest heights would indicate if the crown width upper bound needs to be increased.