Abstract:

Interferometric synthetic aperture radar(InSAR)has been widely used in monitoring natural disaster. SAR image co-registration processing, projecting the secondary image into the same grid of the primary image, is a key step in InSAR data processing prior to the interferogram generation process. Inaccurate image co-registration can significantly reduce the quality of the InSAR interferogram generated. Therefore, an accurate image co-registration result is very important for generating high quality interferograms during the InSAR data processing.
Conventional SAR image co-registration method selects the regular tie points, which are uniformly distributed throughout the SAR intensity image with certain spacing, namely regular points(RP). Since these tie-points are selected without considering the scattering properties of the target, it performs well in region with relatively high coherence. However, most tie-points often fall into the low coherence region, leading to reduction of efficiency, accuracy and reliability. Aiming to deal with the problems of large computation requirement, high mismatch rate and poor co-registration accuracy of conventional co-registration methods, a co-registration method is proposed that combines speeded up robust feature(SURF), external geographic data and cross-correlation. The processing workflow of the proposed method is as follows: First, patch processing is applied to the primary SAR intensity image to separate the image to multiple patches. Second, the amplitude value of the data at each patch is then linearly stretched between 0~255. Third, SURF operator is used to extract the feature points from the primary intensity map. Forth, the external geographic data are used to remove improper feature points and turn them into the remaining feature points for cross-correlation processing. Finally, the co-registration transformation equation between the primary and secondary images is calculated through multiple formulas.
The C-band Sentinel-1A SAR data from two case studies, Pearl River Delta region and Chile earthquake, are analyzed to assess the performance of the proposed method. In this study, the performance of the proposed method is analyzed in various ways: 1)The pixel shift in the azimuth and range direction at the tie-points obtained from the proposed method(i.e. feature points, FP)and the conventional method(RP)can be compared. For the purpose of consistency, we have controlled the tie-point selection criteria such that the amount of FP and RP are close in number. Regions with no or very low pixel shift are first selected. By assuming that no pixel shift occurs in these regions, it is possible to assess the precision of the results obtained from the two methods based on the difference between the observed pixel shift and 0(no shift). We used 0.1 pixel as an indicator for the assessments. It is found that after the removal of the waterbody tie-points, approximately 80% of the tie-points have the pixel shift between -0.1 and 0.1 pixel at the range direction for the proposed method, while the conventional method counterpart is approximately 54%. For the case of the pixel shift at the azimuth direction, the percentages of the tie-point with pixel shift between -0.1 and 0.1 pixel are 82% and 62% for the proposed method and the conventional method, respectively. The result suggested that the pixel shift obtained from the proposed method is more reliable than the conventional method. 2)We then compared the correlation coefficient of RP and FP obtained from the conventional method and the proposed method, respectively. It is found that the peak correlation coefficient obtained from the proposed method is approximately 0.9, which is much better than the conventional method. Moreover, only 43% of the tie-points obtained from the conventional method have correlation coefficient higher than 0.45. the proposed method shows significantly larger portion of the tie-points with correlation coefficient higher than 0.45, which is approximately 94%. 3)The accuracy and efficiency for the transformation equation fitting with the FP and RP are assessed. It is found that, FP only requires 1/4 of the tie-points and computation time compared to the RP for achieving the same expected accuracy. This suggests that the proposed method can significantly improved the efficiency compared to the conventional method. 4)The 2015 M_W8.3 Chile earthquake is used as a case study to compare the performance of the conventional method and proposed method to be used for pixel offset tracking. The results show that approximately 75% of the data obtained from the proposed method can be used to deliver the final displacement field, while only 30% of data can be used for the conventional method counterpart. This suggests that the proposed method has significantly improve the efficiency. The displacement values obtained are compared to the GPS observations, and it is observed that the RMS error in range and azimuth direction is approximately 15cm and 19cm, respectively, which is about 1/15 pixel. the experimental results from the two case studies, Pearl River Delta and Chile earthquake, show that, compared with the standard co-registration method, the method proposed in this study has the advantages of high co-registration efficiency and strong reliability.

Key words: coregistration, Speeded Up Robust Feature, feature points

摘要：

合成孔径雷达(Synthetic Aperture Radar, SAR)影像配准是合成孔径雷达干涉测量(Interferometric SAR, InSAR)数据处理的一个关键步骤, 是生成良好干涉图的必要条件, 但由于各种因素导致很难实现精确配准。针对常规的配准方法计算量大、 误匹配率高以及配准精度差的问题, 文中提出了一种利用特征点进行配准的方法。首先, 使用加速鲁棒特征算子(Speeded Up Robust Feature, SURF)从主强度图中提取特征点; 然后, 结合外部地理数据剔除不当的特征点, 并对剩下的特征点进行互相关处理; 最后, 通过多项式拟合计算主、 辅影像之间的转换模型。通过珠江三角洲地区与智利地震的SAR影像的配准实验结果表明, 文中提出的方法与标准的配准方法相比具有配准效率高、 可靠性强的优点。

关键词: 配准, SURF, 特征点