The filter test has been completed. A paper on the test will be presented at the "3-D reconstruction from airbone laserscanner and InSAR data" workshop in Dresden (8-10 October 2003). We would like to take this opportunity to extend their gratitude to those who participated in the test. Without their input this test would not have been possible.
George Sithole, George Vosselman
This report was compiled while both authors were at the Delft University of Technology, the Netherlands. George Sithole is now at the University of Cape Town, South Africa. George Vosselman is now at ITC, the Netherlands.
Abstract
As one of the tools for rapid topographic feature extraction, the commercial use of airborne laser scanning (ALS) has gained wider acceptance in the last few years as more reliable and accurate systems are developed. While airborne laser scanning systems have come a long way, the choice of appropriate data processing techniques for particular applications is still being researched. The tasks in data processing include the "modeling of systematic errors", "filtering", "feature detection" and "thinning". Of these tasks manual classification (filtering) and quality control pose the greatest challenges, consuming an estimated 60 to 80% of processing time and thus underlining the necessity for research in this area.
Numerous filter algorithms have been developed to date. To determine the performance of filtering algorithms a study was conducted in which eight groups filtered data supplied to them. The study aimed to determine the general performance of filters, the influence of point resolution on filtering and future research directions. To meet the objectives the filtered data was compared against reference data (contained in eight data sets) that was generated by manually filtering real ALS data.
For the purposes of the test, the ALS data was defined as an abstraction of a landscape. This definition was necessary for distinguishing between the Bare Earth and Objects. The landscape was defined as being composed of the Bare Earth and Objects. Objects were further defined as being either Detached (free of the Bare Earth, e.g. buildings) or Attached (connected to the Bare Earth, e.g. bridges).
Having conceptually defined the landscape, seven characteristics of filters were identified based on the filter algorithms submitted and other filter algorithms documented in publications. These characteristics were set aside because it was judged that they influenced the performance of a filter. The seven characteristics are (1) data structure, (2) test neighbourhood, (3) measure of discontinuity, (4) filter concept, (5) single vs. iterative processing, (6) replacement vs. culling, and (7) use of first pulse and reflectance data. These characteristics were used to understand the behaviour of filter algorithms.
Each filter algorithm was then studied, and the result of the output of the filter algorithms was visually compared against reference data. This formed the qualitative comparison. The main problems faced by the filter algorithms were in the reliable filtering of complex scenes, filtering of buildings on slopes, filtering of disconnected terrain (courtyards), and the preservation of discontinuities. Fifteen sub samples were extracted from the eight data sets. The fifteen samples were representative of different environments, but more focused in respect to the expected difficulties (as determined by the qualitative comparison). The output of the filtered algorithms was numerically compared against these fifteen sub samples. This formed the quantitative comparison.
From the results it has been found that in general the filters performed well in landscapes of low complexity. However, complex landscapes as can be found in city areas and discontinuities in the bare earth still pose challenges. It is suggested that future research be directed at heuristic classification of point-clouds (based on external data), quality reporting, improving the efficiency of filter strategies.
Downloading the report
The full report is composed of the main document and two appendices. The report can be viewed in either PDF or HTML format.
Main document ( PDF, 1.4 Mbyte)
Appendix A ( PDF, 0.6 Mbyte)
Appendix B ( PDF, 17 Mbyte!)