The Morayfield Lidar data was verified using differential GPS field survey which demonstrated that the RMS error for vertical accuracy of the Lidar data ranged between 3mm to 3cm. The GPS field survey coordinates was further used for image rectification. Digital cadastral data was available for both the areas. Further Lidar, imagery and cadastral data was made available for Adelaide, South Australia for testing of the algorithm in an urban area with different fence-line characteristics.
L3Harris Geospatial received the data sets for development of automated fence-line extraction techniques at its Broomfield, Colorado headquarters. Atle Borsholm, Senior Software Developer , took the lead in creating separate extraction methodologies for the LiDAR data and imagery, and then comparing the results to determine which worked better.
Borsholm worked in the ENVI image analysis package, an ideal environment for the pilot because it offers a variety of algorithms created for identification, mapping and extraction of specific features in various types of remote sensing data. In addition, ENVI allows the user direct access to the underlying IDL development language used to create the software. This enables the user to customize algorithms and combine them with existing ones to build complex geospatial data analysis workflows.
“The first step was to utilize existing LiDAR point classification tools in ENVI,” said Borsholm.
He explained that some built-in classification algorithms identified buildings, vegetation and terrain. There was no existing algorithm for fences, but the primary attribute of these features – an almost perfectly straight horizontal geometry – helped identify them in the point cloud. ENVI has an existing function for classifying linear features in LiDAR point clouds, which provided a good start for the extraction. The challenge was differentiating fences from other long, straight objects, like kerbs and powerlines.
“We used filters to eliminate these false positives by their height or Z values in the LiDAR data sets,” he said, explaining that any linear feature taller than eight feet was assumed to be a powerline. Features shorter than two feet were likely kerbs or road edges. Likewise, any linear feature wider than six inches was filtered out as non-fence.
The sides of buildings are often also straight lines, but the ENVI classification algorithms had no trouble distinguishing them from fence lines. One source of trouble in the urban/suburban LiDAR data was bushes and other vegetation growing against, and even on top of, fences. Borsholm wrote line fitting algorithms to separate vegetation, which usually has a rough texture in a point cloud, from the smooth fence segments the bushes were obscuring.
“We achieved good results classifying many types of fences – solid wood and chain link,” he said.
Not surprisingly, the denser LiDAR point clouds with 64 ppsm yielded better results in both the rural and suburban/urban applications. In the rural test area where fence lines separating large ranch properties are much longer than in suburban neighborhoods, the automated extraction achieved nearly the same positive results with low-density point clouds as it did with the high-density data set. The lengths of the ranch fences made them extremely easy to identify with the algorithms.
For extraction from the orthorectified imagery, Borsholm created a similar workflow using a combination of built-in and custom algorithms in ENVI. As with LiDAR data, the software had existing tools to automatically classify linear features in digital imagery, which worked well in the pilot project. On the downside, the image data lacked the Z values needed to create filters to eliminate certain features – powerlines and roof edges – by their heights.
“As a result, the automated extraction from the imagery returned more false positives than from the point clouds,” said Borsholm. “The LiDAR technique correctly identified more fences.”