(i) Optical image (ii) SAR image and (iii) Radar spectogram of target
Duration of the Thesis: 6 months
Completion: July 2009
Tutor: Prof. Dr. Mihai Datcu (DLR, Oberpfaffenhofen)
Examiner: Prof. Dr.-Ing. D. Fritsch
Data collected by SAR sensor and processed by SAR processor is inherently complex valued. For the ease of interpretation to user, it is normally provided in detected format, but the phase information is lost during this process. The spectrum of complex valued SAR data which contains this phase information has a special meaning. Thus an attempt has been made to develop and suggest methodologies to expolit this special meaning of spectrum to extract information from complex valued TerraSAR-X images for classification. Two techniques, both based on spectral analysis, have been discussed in this thesis. First technique, based on spectral decomposition, denoted as radar spectrogram [1] & [2] has been found quite suitable for manual target analysis. The second technique, based on short time Fourier transform gives quite encouraging results for classification of the images. Six nonlinear feature arrays have also been proposed which are the input parameters for clustering algorithms. The main objective of this research work is to generate temptations to use complex valued data over detected data and provide a new dimension to the high resolution SAR image understanding.
The goal of signal processing is to extract information embedded in the signals by various possible means. While dealing with complex signals, choosing tractable technique for analysis and information extraction from various possible options or developing a new method is not an easy task. In similar way, the information extraction from image data, a two dimensional signal is also a mammoth task, as analysis is dependent upon various parameters of the image such as sensor, resolution and data format etc.. In case of satellite imagery, the automated method of information retrieval has always been a challenge. Moreover the ever increasing resolution in modern day satellite images has tremendously increased the complexity of problem. With low resolution images, interest was always in global criterion. But the increasing resolution has prompted to analyze targets at finer level. The problem of target analysis in Synthetic Aperture Radar (SAR) is dealt differently as SAR is not an imaging sensor but an active coherent device, which works on the principle of transmitting coherent electromagnetic pulses and recording the amplitude as well as phase information of backscattered signal. Depending upon the requirement SAR data is available in various formats, complex valued as well as detected format. Because of the availability of phase information in complex valued data, it is possible to do detail target analysis based on phase and amplitude information. In this thesis, it has been attempted to exploit the holographic property of SAR signal for the fine backscattering analysis of targets and perform spectral analysis of complex SAR data [3]. A method based on the principle of short time Fourier transform have been proposed for spectral analysis of complex SAR data. The information extracted from this method, and compressed in the form of six non-linear feature vectors acts as input parameter in k-means and SVM based clustering algorithms to obtain classification of image. The proposed feature vectors are motivated from timbral texture features used for music genre classification [4]. Another method based on spectral decomposition denoted as radar spectrogram has also been analyzed in this thesis. TerraSAR-X SpotLight Single Look Complex data have been used as test data set for the mentioned analysis, and obviously the scheme may be extensible to SAR sensors of similar resolution working on different frequency bands.
In this thesis, we have focused on spectral analysis of complex valued SAR data for information extraction at finer level. We have mainly discussed two techniques based on spectral analysis.
First technique based on spectral decomposition, called radar spectrogram was proposed by CNES in [1] & [2] for the analysis of fine backscattering behavior of various targets. Radar spectogram, a time frequency analysis (TFA) approach can be assumed as a generalization of the azimuth splitting method. The cut of the spectrum allows the study of the phase responses of scatterers, which have been viewed by the antenna at different viewing angles. The principle exploits the holographic property of the spectrum at the cost of reducing resolution. As we can see from the presented results in Fig. 1, radar spectogram technique provides a very powerful tool for individual target analysis. Although radar spectogram is very complex in nature, but this complexity can be exploited for use of this method in applications such as change detection.
|
| Fig. 1: Radar spectrogram (i) Optical image (ii) SAR image and (iii) Radar spectogram of target |
The other advantage of high resolution data is that it can be considered stationary when looked in short time. In discussed second techniques, non-linear short time Fourier transform analysis carried out in the thesis try to exploit the local stationarity of the signal. Clustering results obtained from non-linear STFT analysis are found to be very encouraging. The clustering by radar spectrogram was found to be very time consuming. Thus it can be attempted to extract individual target information using radar spectrogram technique, and combining it in the classification results obtained from non-linear STFT technique. Fig. 2 shows the classification result obtained using non-linear STFT technique.
 |
| Fig. 2: K-means based classification using non-linear STFT |
It can be concluded that this research work provide new temptations to use complex valued data over detected data. This research work also provides a starting point in analyzing the high resolution SAR images by using the spectral content. Thus providing a new dimension to the information extraction from high resolution SAR data, leading to a more robust and accurate classification. Further extensions of the work are also possible. For example, now no contextual information is used in the analysis of targets. The inclusion of contextual information can be used to apply this method for content based image retrieval systems. Along with the spectral information, inclusion of textural information in feature vectors is expected to increase the robustness of classification and information extraction capabilities of system.
[1] C. Tison, J. C. Souyris, N. Clerc-Renaud. Point target behavior in high resolution SAR Images: time-frequency versus polarimetric analysis. EUSAR 2006.
[2] M. Spigai, C. Tison, J. Souyris. Time-frequency analysis of point target behavior in high resolution single polarization SAR images. EUSAR 2008.
[3] J. Singh, M. Soccorsi, M. Datcu. Parametric versus non-parametric complex-values image analysis. IEEE International Geoscience and Remote Sensing Symposium, July 2009. (yet to be published).
[4] T. Li, M. Ogihara. Towards Intelligent Music Information retrieval. IEEE Transactions on Multimedia, pages 564-574, June 2006.