Perceptual grouping for automatic detection of man-made structures in high-resolution SAR data
Introduction
The analysis of remote sensing data is a major field of application for pattern recognition methods, e.g. terrain classification and object detection, classification or identification. Extraction of semantic information from the image data is required for a variety of purposes such as land surveying, monitoring of urban growth and natural hazards, disaster management or defence issues. Due to its independence from the time of day and all weather capability, synthetic aperture radar (SAR) has become a key remote sensing technique, complementing sensors working in the visible or infrared frequency domains. The analysis concept for SAR data has to match the features of the studied imagery, e.g. the spatial resolution. In the case of data with coarse spatial resolution (e.g. ERS 1 satellite data of approximately 25 m cell size), the analysis is often restricted to radiometric image properties. A typical statistical pattern recognition task applied to these types of data would be to classify land cover properties.
In the 1990s, airborne SAR sensors achieved spatial resolutions of about 1 m in range and even better in azimuth (Ender, 1998). Some prominent geometric features of urban objects can be identified in such SAR images. Therefore, the focus of the image interpretation shifted from radiometric to geometric properties of man-made objects such as buildings. Model-based structural image analysis approaches for the recognition of urban structures in such data have been proposed, e.g. for reconstruction of buildings (Bolter, 2001, Gamba et al., 2000, Stilla et al., 2001, Soergel et al., 2003) and road networks (Tupin et al., 1998) or the grouping of regular point structures on extended building blocks (Michaelsen et al., 2002). However, only rather coarse levels of detail have been achieved. Now, experimental airborne SAR systems such as the PAMIR system (Ender and Brenner, 2003) can resolve objects of size even below one decimetre. The focus of this paper is to demonstrate new opportunities for structural image analysis, which arise from the visibility of detail structures of urban objects in high-resolution SAR data. A remote sensing application of large interest is building reconstruction for the generation of 3D city models (Gruen et al., 1997). As a preliminary step towards this goal a method for segmentation and grouping of salient structures of buildings is proposed. Concepts of human perception (Wertheimer, 1923) are integrated in a knowledge-based system (Stilla et al., 1996) for the automatic detection of man-made structures. In contrast to complex high level reasoning systems such as (Matsuyama and Hwang, 1990) the presented system is rather simple and contains only a few production rules. However, it is capable of processing large amounts of data. Most benefit from such a technique is expected in an intermediate process level, i.e. following the feature extraction but preceding final decisions or automated high level reasoning.
Section snippets
Buildings in high-resolution SAR data
In Fig. 1a and b two SAR images of the same scene (150 m × 150 m) of 1 m resolution (Fig. 1a) and 20 cm resolution (Fig. 1b) respectively are shown. Both SAR acquisitions have been carried out from approximately the same aspect and distance. Comparing the two SAR images, it is obvious that buildings in particular look very different. One reason for this is the larger dynamic range of the high-resolution data (in this case about 70 dB). Eaves and ridges of the roofs can be identified clearly as linear
Declarative knowledge for non-local building recognition
Many building recognition approaches directly classify from the pixel-data or from local features such as wavelets. A good investigation on what can be achieved by such methods today is found in Bellman and Shortis, 2004. For such approaches only a limited image area at limited resolution can contribute to the decision. It is obvious that in high-resolution SAR data some of the most distinctive properties of man-made structure in an urban area are of non-local type. Their understanding requires
Perceptual grouping using a production system
First steps in the workflow of a production system are primitive object segmentations. Based on these the search starts by applying production rules that build step by step more complex objects, which are added into the data base of accumulated objects. In the following sections the production rules, the control scheme, and the feature segmentations are discussed. Additionally, we give some comparison with possible standard state-of-the-art solutions in Section 4.3. An overview of the whole
Results
The high-resolution X-band SAR image presented in Fig. 1b has been processed using the production system and data driven control as outlined in the previous sections. The system is implemented in machine independent Matlab code. Table 2 displays the growth of the numbers of objects during the search run. Fig. 4 shows the progress after 40,000 control cycles. The search using the proposed production system returns a set of reduced objects. There may exist many local alternative objects among
Conclusion
State-of-the-art high-resolution SAR sensors provide a detailed mapping of man-made objects, which could not be achieved by radar remote sensing only a few years ago. It was shown that this offers new opportunities for structural image analysis approaches in urban scenes, e.g. for building reconstruction. Particularly, a much finer level of detail of object recognition seems to be possible. In order to address the enormous amounts of data provided, multi-scale and parallel processing are
Acknowledgements
We want to thank Prof. Dr. Ender and Dr. Brenner (both FGAN-FHR) for providing the SAR image data. The data were recorded by the experimental systems AER-II and PAMIR of FGAN.
References (22)
- et al.
Extracting contours by perceptual grouping
Image Vision Comput.
(1992) Which Hough transform?
CVGIP; Image Understanding
(1993)- et al.
An optimizing line finder using a Hough transform algorithm
CVIU
(1997) - et al.
A classification approach to finding buildings in large scale aerial photographs
Int. Arch. Photogrammetry Remote Sensing
(2004) - Bolter, R., 2001. Buildings from SAR: detection and reconstruction of buildings from multiple view high resolution...
Experimental results achieved with the airborne multi-channel SAR system AER-II
Proc. EUSAR
(1998)- et al.
PAMIR—a wideband phased array SAR/MTI system
IEE Proc.—Radar, Sonar, Navigation
(2003) A framework for low level feature extraction
- et al.
Detection and extraction of buildings from interferometric SAR data
IEEE Trans. Geosci. Remote Sensing
(2000)