Where Are We With Feature Extraction?

Feature extraction and automated feature extraction were the main topics of discussion at conferences and during presentations about five years ago. Now, focal plane LiDAR and UAV’s are all the buzz. Feature extraction is still a vital part of what we do and are trying to do better. Where are we with feature extraction and how have we progressed?

Companies are very proprietary about their extraction processes so to find out where we are as an industry is very difficult. The big measure of this is from the software companies and one thing is clear, the mobile LiDAR profession has come a long way in developing automated processes to extract features although some processes are still manual. Orbit GT”s Geospatial and Certainty 3D’s TopoDOT have done a great job developing automated feature extraction processes for an array of features. The success of the mobile LiDAR automated feature extraction relies on the density of the points and the geometry associated with those points.

There are several LiDAR software companies that have some variation of automated feature extraction for airborne LiDAR including, but not limited to Merrick MARS®, Harris ENVI, Trimble eCognition, TerraSolid, GeoCue LP360 and several others that do automated feature extraction from airborne LiDAR. These software packages have variations of what they exactly do as it relates to automated feature extraction.

Currently, Merrick & Co. MARS® software provides some automation, but the amount of automation required to provide a competitive solution for the transportation application has not been fully realized. Merrick & Co. provides the following examples of fully automated processes with varying results based on features extracted; Rough roof outlines, roof elevation heights, rough 3-D building wire frames, vegetation polygons, powerline, attachment points, and rail extraction.

It should be asked is there any way to apply what is being done for feature extraction with mobile LiDAR to aerial LiDAR. Perhaps this would work with higher density helicopter data, but it probably can’t be achieved with fixed wing LiDAR, or more specifically data of less than approximately 25 points per square meter (PPM).

orbitMerrick & Co. is currently working with Orbit GT to do a feasibility and viability assessment of the Orbit software as it relates to semi-automated and automated feature extraction from high density airborne LiDAR and high resolution airborne imagery. The goal of this exercise is to apply techniques used in extracting features and apply the same analytics used in mobile data, to airborne data. The results will be reviewed and compared and then modification to the analytics will be applied to see if the results can be improved.

To insure accurate results a complete mobile and airborne data set is being provided to Orbit GT of the same area. The mobile data was collected for a section of highway and the same section of highway was collected with a helicopter-based, high density mapping system (HDMS). The HDMS system was configured to collect LiDAR at 25 to 35ppm and the digital imagery was configured to collect 2-inch pixel resolution. The airborne and mobile data was referenced and co-registered to the same control. The data was collected for 1:30 scale mapping.

The data for this project was collected and compiled for a significant transportation and highway safety project. The data deliverables are LiDAR, imagery, and a geodatabase. The detailed transportation plan data is going to be used to assess safety issues and concerns throughout a highway network. The intent of the data is to make driving safer and reconstruct problem locations to make them safer for drivers and reduce or eliminate accidents at these locations.

Plan&imageryThe test data set was processed using the airborne data so that the results could be provided to Orbit GT for comparison. The feature extraction was manual and very time consuming. The sample data set included the mobile data for the same area so that feature extraction could be run on both the mobile data and the airborne data. The results will give an indication of if and how the algorithms can be modified based on the airborne data’s point density and imagery.

It stands to reason, based on all of the automated feature extraction tools currently available that it should be possible to achieve a solution for this application. Once the assessment is completed additional information will be provided.



James Wilder Young (Jamie) CP, CMS-L, GISP is currently a Senior Geomatics Technologist for Merrick & Co. located in Greenwood Village, Colorado. His experience includes all aspects of LiDAR including sensor development, applications development, data acquisition, data processing and project management.

Contact Info:

Jamie Young –  jamie.young@merrick.com

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.