Our research project is about segmentation and classification of remotely sensed LIght Detection And Ranging (LIDAR) data. Airborne laser scanned LIDAR is a relatively young medium in the field of photogrammetry. The accuracy in elevation of remotely sensed LIDAR data is a beneficial feature for various applications such as forestry, archaeology, 3D city map generation, flood simulations, coastal erosion monitoring, landslide prediction, corridor mapping and wave propagation models for mobile telecommunication networks and many more.
Contributing to these applications, our work aims to develop novel algorithms for terrain feature classification based on LIDAR data. Terrain features such as buildings, streets, railways and bridges, but also rivers, basins and vegetation are subject to our investigations. By using pattern recognition techniques, methods of photogrammetry and the unique height information of LIDAR data, segmentation and classification of terrain features is carried out in the course of the project. Our current investigations aim at terrain feature classification in LIDAR data by fusing several simultaneously recorded bands such as
The measurement technique using LIght Detecting And Ranging (LIDAR) was developed by the National Aeronautics and Space Administration (NASA) in the 1970s and has been commercially used since the early 1990s. A LIDAR data acquisition system consists of three elements: a Laser Range Finder (LRF), a Global Positioning System (GPS) receiver and an Inertial Navigation System (INS). While mounted on a platform like a plane as depicted in Figure 1, the distance between the instrument and a point on the surface is estimated (Z component) by measuring the time the laser pulse needs to hit the receiver of the laser instrument. GPS and INS complement the data sets with position components (X and Y) and orientation, respectively. Since only one measurement per X and Y component pair is estimated, LIDAR data is often referred as 2½ dimensional. Characteristic for LIDAR data is also that their elements are irregularly distributed in the first place and thus form a point cloud. Overlapping strips are usually acquired, when scanning the target landscape. These strips are then combined in post-processing steps, also known as strip adjustment.
Figure 1: Airborne LIDAR data acquisition
Generally, there are at least two types of echoes which can be recorded by a LIDAR system: the first and the last echo as schematically depicted in Figure 1. These two categories of responses are the basis for retrieving Digital Surface Models (DSM) and Digital Terrain Models (DTM). Acquired points of the category of first echoes mostly hit canopies, roofs or chimneys, whereas last echoes of the laser are reflections of the ground. Simultaneously, intensity data, that is the reflected signal power by the objects in the scanned area, can be recorded. This type of data is especially useful if different objects have the same height but different emission characteristics (e.g. water and road). Together with LIDAR intensity (reflectance) and multi spectral data recorded at the same time, segmentation and classification techniques can be exploited extensively. This complementary process of LIDAR data is also referred as Data Fusion.
Figure 2: Pseudo coloured height information of one tile |
Figure 3: Pseudo coloured gridded IDAR data tile |
Figure 2 shows the pseudo coloured height information of LIDAR data supplied by Environment Agency, UK. Noteworthy is the noise that generally comes with LIDAR data which makes the development of segmentation and classification algorithm challenging as well as the artefact "no data" as depicted in the lower right corner in Figure 3. Table 1 summarises both the advantageous but also less beneficial characteristics that make LIDAR special:
| Pro | Con |
|---|---|
| no shadows | no surface texture, no colour information |
| no horizontal occlusion | vertical occlusion |
| high accuracy for less hilly terrain (10 cm - 15 cm) | poor accuracy in hilly terrain (up to 200 cm) |
The most common problem in LIDAR filtering is the separation of object and ground points resulting in DTM and nDSM generation. On principle, two approaches can be chosen to segment and classify LIDAR data: unsupervised classification (often referred as segmentation) and supervised classification (also classification). Regarding land cover classification, current investigations target employing Maximum Likelihood Classification. With respect to segmentation, our research involves textural analysis using Wavelets, Gabor Wavelets, modified histogram thresholding using Wavelet Packets and co-occurrence matrices. We also investigate on the separation of object and ground points in LIDAR data (irregular and gridded) using Skewness Balancing. For further information, please refer to our publications.
Figure 4a illustrates the generation of a Digital Terrain Model (DTM) from a Digital Surface Model (DSM) for an artificial height scene characterised by sloped terrain and several buildings in Figure 4a using Skewness Balancing. The normalised Digital Surface Model (nDSM) in Figure 4b is obtained by subtracting the DTM from the DSM.
Figure 4: Generation of a DTM and an nDSM from a DSM
Figure 5 depicts a LIDAR data tile of an urban area with detached objects (both buildings and vegetation of different height) and the scene after applying Skewness Balancing. The data was kindly provided by TopoSys GmbH, Germany, by courtesy of the Stadt Mannheim, Germany, the copyright holder ©. It can clearly be seen that all detached object have been removed. Further details can be found in our paper Segmentation of LIDAR Data using Measures of Distribution, presented at the ISPRS Mid-term Symposium 2006 "Remote Sensing: From Pixels to Processes", Enschede, the Netherlands, 8-11 May 2006.
Figure 5: Object and ground point in LIDAR data separation using Skewness Balancing
Poster Paper Prize - Merit Award - at the RSPSoc 2006 at Fitzwilliam College, University of Cambridge, for the paper
M. Bartels and H. Wei.
Rule-based Improvement of Maximum Likelihood Classified LIDAR Data Fused with Co-Registered Bands.
Annual Conference of the Remote Sensing and Photogrammetry Society, CD Proceedings, 2006
Cambridge, UK, 05-08 September 2006
|
Address: |
|
Marc Bartels |
|
|
Hong Wei |
|
Hong Wei and Marc Bartels at the ICPR 2007. |
We would like to take the opportunity to thank our supporters to this project. The project is RETF funded by the University of Reading and has been supported with conference and travel grants by the School of Systems Engineering, the University of Reading, the RSPSoc, UK and IAPR. We would like to thank the data providers for LIDAR data supply, in particular
