Planners are increasingly looking at ways to increase tree canopy in order to deal with a myriad of issues from reducing the urban heat island effect to improving wildlife habitat. In early 2014 the Northern Kentucky Urban & Community Forestry Council (NKUCFC) commissioned an assessment of the tree canopy in Boone, Campbell, and Kenton counties in order to inventory the existing tree canopy and prioritize locations for planting new trees. The SavATRee Consulting Group, in collaboration with the University of Vermont Spatial Analysis Laboratory, was given the task of carrying out the assessment.
The foundation of the assessment was a high-resolution (1.5 foot) land cover dataset that would map the regions gray (buildings, roads, and other paved surfaces), green (tree canopy and grass/shrubs), blue (water) and brown (bare soil) infrastructure. With over 37,000 acres to map in the tri-county region the resulting land cover dataset was set to exceed 11 billion pixels in size. This coupled with an aggressive timeline and a mandate map tree canopy with a minimum of 94% accuracy made for a challenging project. All three counties had robust GIS datasets that accounted for their gray infrastructure, but they lacked any detailed vegetation maps. High-resolution leaf-on imagery was acquired for the entirety of the tri-county region as part of the National Agricultural Imagery Program (NAIP) in 2012, but the broad acquisition window and heterogeneous landscape made the imagery, when used alone, less than ideal for tree canopy mapping.
Fortunately, Kentucky has a robust LiDAR program, known as Kentucky from Above. LiDAR data were acquired from the tri-county region in 2011/2012. Although the LiDAR data had been around for nearly two years and other urban tree canopy assessments had been conducted in the state, this marked the first time the LiDAR would be used for such a task. Trees that were either undetectable due to their size, obscured by shadows, or confused with shrubs in the imagery, were clearly visible in the LiDAR. There was one drawback of the LiDAR it was acquired under leaf-off conditions. As a result, large canopy gaps appeared in forested areas dominated by deciduous trees. This was due to the fact that deciduous trees in forested areas have no leaves on in the winter months and they tend to have thin branches as most of the woody growth is centered on increasing the height of the tree to outcompete neighboring trees. These areas were clearly discernable as forest in the imagery.
Given that manual interpretation for such a large area would be both too time consuming and costly an automated approach was needed. The solution for mapping tree canopy required a series of rules that would adjust based on the landscape setting so as to maximize the strengths and minimize the weaknesses of each dataset. Another issue of concern was that the imagery and the LiDAR did not align precisely, with trees in the imagery exhibiting some lean as they were not true orthophotos.
The automated land cover mapping system was built on Trimbles eCognition platform. eCognition is an objected-based feature extraction software capable of working with raster, point cloud, and vector datasets. The starting point for feature extraction was the LiDAR point cloud in LAS format. A batch routine was employed to load the 1,476 LAS files into eCognition. The same rule-based expert system was then applied to each LAS tile.
Figure 1 Example of the final 7-class land cover product.
The expert system consisted of three main parts: 1) data loading, 2) feature extraction, and 3) data export. In the data loading phase the image mosaic and various raster LiDAR models derived from the LAS files were loaded, along with the available vector layers. As eCognition virtually subsets each loaded layer to the base dataset (the LAS file) no data clipping was required. In the feature extraction phase a series of segmentation, classification, and morphology algorithms were used to extract the land cover features. In those cases where a land cover class had already been mapped (e.g. buildings) that class was simply incorporated.
For the tree canopy, a series of contextual rules was used to adjust the influence of the LiDAR and imagery datasets depending on the landscape. The use of LiDAR was emphasized for street trees, which had distinct profiles in the LiDAR, but were barely visible in the imagery. For the aforementioned deciduous forests, a complex iterative process was developed that identified these forested stands using a combination of the LiDAR and the imagery.
The resulting land cover dataset was a huge success. Trees were mapped with 99% accuracy, exceeding the project goals and surpassing the accuracy achieved from previous studies. Without a doubt the project could not have been accomplished without the LiDAR data from Kentucky from Above. Tree canopy mapping was not likely even a consideration when the state contemplated LiDAR, proving that new uses will always be found for this technology.
Figure 2 Example of a canopy height model for the region developed from the land cover and LiDAR datasets.