While automated computer vision has proven very successful in well-constrained industrial environments where illumination, object types, and orientations are known a priori, it is exponentially more difficult to analyze and extract features in high-resolution satellite and/or airborne earth imagery given the wide intra-class feature variations (spectral, spatial, context, etc.) found within individual scenes and spatio-temporal variations across many scenes. However, given the rapid growth in computational capacity and the corresponding development of more sophisticated image processing techniques, we believe that automated feature extraction (AFE) algorithms can be developed to achieve a level of proficiency required for robust upstream processing of multi-modal datasets.

Much of our AFE research has focused on urban-area feature extraction utilizing high resolution Commercial Satellite Imagery (CSI) like that shown below from Space Imaging's Ikonos satellite. Automated extraction of urban features and fusion of these features with other intelligence information can produce "smart" vector products that contain a wide variety of information for each feature. CSI data typically have four multi-spectral (MS) bands (R, G, B, NIR) and one panchromatic (PAN) band with spatial resolutions on the order of 2.4-4 m for the MS data and 0.6-1.0 m for the PAN data.

Advantages of CSI data are as follows:

• Spatial resolution is comparable to some imagery from National Technical Means (NTM)
• It is easily shared with researchers, state & local governments, and U.S. allies
• Unlike NTM, it can be used within the U.S. borders for homeland security needs
• It is an integral part of U.S. government source data through the $500M NextView contract

Road Extraction

We have developed a number of fully automated road extraction algorithms and additional research in this area is ongoing. A linear feature extraction (LFE) algorithm is illustrated below where the directional spectral similarity is measured for each non-vegetated pixel (identified using an NDVI threshold). Large linear segments with a high length-to-width ratio are identified as candidate road segments. These are then combined in a piecewise linear fashion to form road networks even when the network topology contains curvilinear roads.

Another approach utilizes image segmentation followed by perceptual grouping rules to combine both segments and road centerlines derived from the segments. An illustration of this is shown below.

We then employ multi-detector fusion strategies to combine the output of the different road detectors in an optimal way. The approach dynamically and automatically uses the road detector output that is best suited for a particular road network topology (gridlike, curvilinear, etc.) based upon local image analysis. Examples of the integrated multi-detector road extraction results for an urban and suburban area are shown below.

Evaluations of the extracted road network using representative test sites show completeness values that range between 70-86% and correctness values that range between 70-92%. These are amongst the highest accuracies for fully automated road extraction reported in the literature thus far.

 

Building Extraction

We have developed a number of fully automated building extraction algorithms and additional research in this area is ongoing. Multi-detector strategies are needed for robust performance given the heterogeneous nature of the spectral signatures, structural morphology, and spatial topology of building features in high-resolution satellite imagery.

By applying mathematical morphology operators to high-resolution images, the multi-scale image structure can be decomposed. Morphological operators can separately identify objects/structures that are brighter or darker than the surrounding area. The images below show where compact objects, both bright and dark, have been identified in the urban image at one particular scale. Many of the bright building structures in the image are captured in the center image, and many of the dark L-shaped building shadows are identified in the right image. This type of decomposition can be performed at many scales to identify features/objects of varying shapes and sizes.

The multi-scale objects can then be analyzed based on shape, size, spectral content, spatial context, etc. to identify and extract 2D building polygons. For example, detection of building shadows can be used to identify the buildings causing the shadows as shown here:

An example of the multi-detector building extraction for a complex dense urban scene is shown below. Automated 2D building vector extraction accuracy for this urban scene was about 73%. This is the highest accuracy for fully automated building extraction reported in the literature thus far.