Home Articles Creation of a 3D Urban GIS Database: Data Fusion ApproachTechnical Session...

Creation of a 3D Urban GIS Database: Data Fusion ApproachTechnical Session on “Photogrammetry and 3D Visualization”

Don Wicks and Abdul Rauf Campos-Marquetti
Spectrum Mapping, LLC
1560 Broadway, Suite 2000
Denver, CO 80202
Email: [email protected], [email protected]

ABSTRACT
This project was centered on the development of a 3D Urban GIS Database for the U.S. Army Research Development Engineering Command that would be used for the purposes of mapping and simulation-visualization. Spectrum’s role was to physically collect (using its in-house LIDAR, Digital Camera and hyperspectral sensors), and develop the necessary source data required to construct a 3D urban database using real-time urban data. This data was developed using a data fusion approach in which LIDAR, Color-CIR Digital Camera and 63-band Hyperspectral data were integrated in order to acquire 3D urban feature data. This base terrain data and its associated feature objects were used as the input into the GIS and visualization models.

LIDAR data served as the terrain and (X,Y,Z) feature source information for most of the 3D Objects contained in the terrain database. This included bare-earth surface, building footprints, height and structure of vegetation and tree cover, roads, and localized sign and street light infrastructure. All features were extracted as LIDAR point data and then transformed into their appropriate terrain formats (GeoTIFF, polygon, point, line shape files). The 0.5-ft to 1.0-ft resolution Color / CIR Digital Orthophotography served as the geo-coordinate base for the terrain database, due to its highly accurate geopositional characteristics, and was used as the base (RGB) drape-overlay layer for the visualization model. The hyperspectral imagery served as the source platform for feature attribution that will be derived using automated spectral analysis techniques (spectral curves-matching and feature space mapping techniques). All hyperspectral pixels were geolocated to their corresponding LIDAR point data, giving each point an identifiable material class to be used in the visualization construction process. The hyperspectral data was also used to generate an overall land cover map, which will provide feature class names and material class attribute for all surface features contained within the project area. Extracted features included: spectral characteristics of buildings, roads; water bodies; water networks; wetlands; agricultural classes, and tree types.

Two primary study areas were used in the project: Los Lunas, New Mexico and Commerce City, Denver Colorado. All resultant data, images and terrain models were input into the ArcGIS version 9.0 environment for display, query and product output.

As part of the work with RDECOM (U.S. Army Research and Development Command), Spectrum generated two project study areas from which earth surface features were to be extracted for use in a prototype urban visualization database. The study areas sued were as follows:

  • Los Lunas, New Mexico Urban-Rural test site
  • Commerce City-Denver, Colorado test site

Each test site was selected for its unique urban characteristics, where urban features were to be extracted using a fusion sensor and dataset approach. The sensors used in this study included:

  • Spectrum RAMS LIDAR
  • RAMS Color Digital Camera
  • SPECTIR Hyperspectral Imager

From the above sensor Spectrum provided a prototype urban database that could be imported into the Terrex visualization software with its associated descriptive database attributes. Each test site was run through the same process for data preprocess, feature extraction and analysis and creation of product deliverables. Several lessons learned were also established that primarily had to do with the feature extraction of residential homes and the characterization in hyperspectral feature space of urban rooftops

Los Lunas, NM Data Acquisition and Processing
The Los Lunas, NM test area included the simultaneous acquisition of high-resolution LIDAR data using Spectrum’s D3 LIDAR system, which provided 1.0-meter spacing data, the acquisition of 0.5-foot pixel Digital Color IR imagery and the collection of hyperspectral imagery using the SPECTIR hyperspectral system. All data was preprocessed to real world coordinates using the on board airborne GPS and IMU data (New Mexico State Plane coordinates NAD83, feet Central Zone). All LIDAR data was preprocessed into a LAS file format, with the Multispectral Digital Color IR camera imagery being delivered as orthorectifed TIFF file formatted files. The hyperspectral imagery was captured as a binary file that was later converted to a georeferenced ERDAS imagine file that contained 63 bands of image data.

The Los Lunas study area is characterized as a rural-urban zone that is dominated by agricultural fields, urban-residential subdivisions, the Rio Grande Valley and associated riverine morphology and riparian forest vegetation, figure 1.

Figure 1. Los Lunas New Mexico Color IR Digital Camera imagery.
Because of the nature of the vegetation in this project area Spectrum chose to fly Color IR digital camera imagery in order to best distinguish vegetation features from the background soil and urban features. The imagery was collected a pixel resolution of 0.5-ft, and provided and excellent cartographic base from which to tie in both the LIDAR and hyperspectral imagery. The imagery was orthophoto processed using the LIDAR data as the DEM base and laid out in a project-tiling scheme and with the resultant 8 orthophotos being mosaiced to form the final project ortho base seen in figure 1.

Once the ortho base was generated feature extraction was initiated using the LIDAR data. This included the extraction of the following earth surface features:

  • Bare Earth surface generation
  • Building footprint generation
  • Vegetation canopy generation

The LIDAR data resolution was 1.0-meter point spacing. Features were extracted using Spectrum’s LID-MAS software using a TIN and Fast Fourier Filter (FFT) in combination. The resultant data included extracted bare-earth surface, buildings and vegetation (trees) in a LAS elevation point cloud format, figures 2.

The LAS extracted elevation point cloud features were then converted to their appropriate feature formats in which they would be delivered. The bare earth surface elevation points were first converted to an ESRI GRID format using an IDW interpolation algorithm (figure 3), and then exported into the ARA deliverable format, which is a 32-bit GeoTIFF file. Buildings were converted from the LAS point cloud format into an ArcView Shapefile as a polygon footprint. Each building polygon in this dataset was assigned the following attributes that included: Area, Perimeter, max building height and min building height.

Figure 2. 3D Perspective of extracted bare-earth surface, buildings and tree point clouds

Figure 3. (Left) Bare Earth Surface-ESRI GRID, (Right) Bare Earth Surface-32bit GeoTIFF, with both images being colored by increasing elevation values.
LIDAR did an excellent job in defining and capturing tree canopies. Tree Canopy point clouds were converted to ArcView Shape files in three distinct formats:

  • Individual Tree Points
  • Individual Trees as polygons
  • Tree Clusters (Riparian Forest) as Polygons

The hyperspectral imagery was used to define and classify the following classes of features, figure 4:

  • Water bodies (river, irrigation ditches, ponds)
  • Non-tree vegetation (Agriculture and native vegetation)
  • Soils: Sandy Loam, Sandy soils, Wet-Clay Soils
  • Road Pavement (Asphalt)

Figure 4. Final Los Lunas Area extracted Land Cover Features. This includes LIDAR extracted buildings, and tree cover; and Hyperspectral extracted, non-tree vegetation cover, agriculture, Soils, road-asphalt and water classes.

Commerce City-Denver Acquisition and Processing
The Commerce City Denver study area included the simultaneous acquisition of high-resolution LIDAR data using Spectrum’s RAMS LIDAR system, which provided 1.0-meter resolution point spacing data, the acquisition of 0.5-foot pixel Digital Color imagery and the collection of hyperspectral imagery using the SPECTIR hyperspectral system. The Commerce City area is characterized as an industrial urban zone and is dominated by industrial warehouses and structures, paved urban streets and highways, and to the west a large residential subdivision. Vegetation cover is strictly urban in nature, primarily dominated by individual trees and some small clusters of riparian tree cover along the riverine area.

The Color Digital Camera imagery was collected a pixel resolution of 0.5-ft, and was used as the cartographic base from which both the LIDAR and hyperspectral imagery were tied in terms of their base coordinate projection system. The digital color imagery was orthophoto processed using the LIDAR data as the DEM base and laid out in a project-tiling scheme and with the resultant 9 orthophotos being mosaiced to form the final project ortho base.

Once the ortho base was generated feature extraction was initiated using the RAMS LIDAR data, which has a spatial resolution of 1.0-meter point spacing. This included the extraction of the following earth surface features:

  • Bare Earth surface generation
  • Building footprint generation
  • Vegetation canopy generation

Features were extracted using Spectrum’s LID-MAS software using a TIN and Fast Fourier Filter (FFT) in combination. The resultant data included extracted bare-earth surface, buildings and vegetation (trees) in a LAS elevation point cloud format, figure 5. In this case most buildings were clearly well defined due to their size and rectilinear morphology. Vegetation cover was relatively sparse with little overhand existing on buildings with the exception of the residential areas.

Figure 5. (Left) Raw Reflectance LIDAR, (right) TIN/FFT filtered extracted bare-earth surface, buildings and vegetation elevation point clouds.
The LAS extracted elevation point cloud features were then converted to their appropriate feature formats in which they would be delivered. The bare earth surface elevation points were first converted to an ESRI GRID format using an IDW interpolation algorithm and then exported into the ARA deliverable format, which is a 32-bit GeoTIFF file.

LIDAR did an excellent job in defining and capturing tree canopies. Tree canopies were relatively sparse in this urban environment and mainly confined to residential and riparian areas. Tree Canopy point clouds were converted to ArcView Shape files in three distinct formats:

  • Individual Tree Points
  • Individual Trees as polygons
  • Tree Clusters (Riparian Forest) as Polygons

The hyperspectral imagery was used to define and classify the following classes of features:

  • Water bodies (rive and man-made ponds)
  • Building rooftop Material
  • Road Pavement (Asphalt)

The hyperspectral imagery used was captured using a SPECTIR 63 band system that captured imagery at a pixel resolution of 1.0-meters. Samples of this imagery can be reviewed in figure 11. Features were classified using image-processing techniques in the form of a semi-controlled unsupervised classification. All classes were verified and examined in detail using the derived class statistics and associated continuous spectral curves. All hyperspectral-classified data was converted from a raster classification to ArcView shape files. A semi-controlled unsupervised classification was then generated that produced three main roof type materials recognizable in that area, figure 6 and 7:

  • Asphalt-tar
  • Gravel-tar
  • Metallic Roofs

All extracted feature data was converted to an ArcView Shapefile format and attributed accordingly as dictated by ARA for use in the Terrex Software system. Figure 8 shows the classification as a 3D perspective.

In Summary the fusion of high-resolution Digital Camera Imagery, LIDAR and hyperspectral data proved to be highly effective in mapping both urban and rural areas. Such data allowed the user to develop land cover classification maps that included a highly accurate cartographic base, a digital terrain model and a material classification map. Such processing techniques allowed the ready development of a 3D urban and rural database with associated feature classes, material composition attribution, linked to highly accurate x,y positional and elevation information at the pixel level of 1.0-meter resolution.

Figure 6. (left) Color Digital Camera data of building sample area. (right) LIDAR building footprint masked and extracted hyperspectral imagery of building rooftops.

Figure 7. Hyperspectral building rooftop classification (left). (Right) Continuous spectral curves of rooftops showing spectral endmembers (Metallic – Asphalt).

Figure 8. Final Commerce City Area extracted Land Cover Features. This includes LIDAR extracted buildings, and tree cover; and Hyperspectral extracted, non-tree vegetation cover, road-asphalt and water classes and digitized road and railroad centerlines.