A Study On Multi-bands And Multi-polarimetric SAR Image Fusion Interpretation

The Synthetic Aperture Radar (SAR) has been widely used in military and civilian areas, because of its weather-independent, all time and strong penetrating capability. However, the radar backscattering information is not only related to terrain types, but also influenced by other factors such as radar observation angle, terrain complex dielectric constant, incident wavelength, polarization etc. These factors lead to the limited capability of obtaining information for single-band single-polarization SAR. With the breakthrough of hardware technology, SAR system can image under multi-bands and multi-polarization modes simultaneously. Therefore, the study of multichannel SAR image fusion has attracted more and more attention. Terrain classification and target detection are two typical applications of SAR image interpretation. Under the support of "Eleventh Five-Year" National Defense pre-research project, this paper focuses on how to use multiple sources SAR images to improve the terrain classification and target detection accuracy. The main work and contribution are as follows:1. The statistical model of background clutter and parameters estimation is studied, which forms the basis of SAR image analysis. Among the existing polarimetric SAR models, G p0 is a good model for characterizing polarimetric SAR images containing different terrain types. Nevertheless, parameters estimation hasn't been well resolved. To solve this problem, this paper proposed a new method based on Log-det cumulant (LDC), experiments on real SAR data demonstrate its superiority over the maximum likehood method (ML) and moment based method (MoM) in terms of speed and accuracy. This laid the foundation for SAR image analysis such as classification and target detection.2. Multiband SAR terrain classification and decision level fusion are studied. Different bands data have different polarization properties, requiring different classfication methods. For single polarimetric SAR data, the classical SVM method is adopted. For fully polarimetric SAR data, by introducing more suitable scattering model and better clutter distribution, we proposed an improved algorithm for polarimetric SAR image classification and obtained better results than Lee's. Moreover, the experiment results indicate that different band data are complementary in distinguishing terrain, so fusion of the classification results at desicion level is realized under the Dempster-Shafer evidence theory framework. Experiments show that the integration of multiple bands information could get better classification results than single band. To evaluate the performance of classification algorithm, cooresponding Google earth image is used as reference instead of real thematic mapping information which is difficult to obtain. Due to the use of entire scene sample data, this method appears more objective than the traditional method of artificial selection testing samples.3. Under the support of National Defence project, this paper takes the lead in camouflage covered vehicle detection. After indepth study of scattering power by covered vehicles under different bands and polarization, we found that the echo power is stronger by longer incident wavelength and more visible by co-polarization wave than cross-polarization wave. Constant False Alarm Rate (CFAR) target detection algorithm is one of the most potential methods. For this algorithm, the background clutter modelling is a key factor in influencing the algorithm performance. A modified global CFAR algorithm is proposed by adopting the improved clutter model and is applied to veiled and unveiled vehicles detection under different band and polarization modes. Detection results are fusioned at decision level according to Neyman- Pearson criterion, which show that the camouflage net could reduces the target detection rate to some extent. Fusion of different band and polarization detection results at decision level could improve the accuracy, reduce false alarm rate and raise the overall detection performance. Moreover, detection performance is raised higher for camouflage covered vehicles than exposed vehicles, which means that multi-band multi-polarization detection fusion is more meaningful for veiled target detection.4. Based on the above proposed classification and target detection algorithms, a typical military scene (airport) is processed and the preliminary interpretation results are obtained, which benefits further battlefield surveillance and situation assessment. Finally, these algorithm interfaces in our software platform are presented.