Minimum-Error Thresholding for Unsupervised Change Detection in Multilook Polarimetric SAR Images

In this paper, we propose an unsupervised method for change detection in polarimetric synthetic aperture radar (PolSAR) images. The symmetric revised Wishart (SRW) is applied for measuring the similarity of two multilook complex (MLC) covariance matrices. The SRW produces a scalar feature image as an input into the automatic thresholding algorithm which is aimed to distinguish change from no change. In particular, the Kittler and Illingworth minimum-error thresholding method is generalized to model the non-Gaussian distribution of change and no change classes. Experimental results on bi-temporal simulated and RADARSAT-2 C-band PolSAR data confirm the effectiveness of the proposed method. The results of the real data also demonstrate that the multipolarization SAR improves the detection accuracy and lowers the overall error rate of the method compared to single-polarisation SAR.
Multitemporal satellite remotely sensed data from a geographical area offers a great potential for monitoring and detecting changes in Earth’s surface (for e.g. damage assessment in natural disasters [1], monitoring the changes in agricultural areas [2], and glacier change detection [3]). Synthetic aperture radar (SAR) is an important instrument in remote sensing, providing measurements insensitive to the sun-light and atmospheric conditions. Furthermore, polarimetric synthetic aperture radar (PolSAR) sensors provide data with increased discrimination capability as compared to single-channel SAR and also insensitive to the sun-light and atmospheric conditions.
Several change detection algorithms in SAR data has been developed in the literature, e.g., [4]-[5]. Unsupervised change detection is generally performed in three steps: 1. image preprocessing including co-registration, speckle filtering, and radiometric and geometric terrain corrections, 2. comparing SAR image pairs with a desired test statistic, and 3. finally, a thresholding method is applied to the test statistic to achieve the final change map.
Fig. 1: the block diagram of the proposed change detection approach
In the analysis of change detection in multilook complex (MLC) PolSAR images, the backscattered signal is represented by the so-called polarimetric sample covariance (or coherency) matrix. In [2], the Wishart likelihood ratio test is proposed as a test statistic for change detection in multilook PolSAR images. Akbari et al. utilized the complex kind Hotelling-Lawley trace statistic as a new test statistic for change detection in multilook PolSAR images [18]-[19]. In [20], Ghanbari et al. applied the symmetric revised Wishart (SRW) distance in [8] as a test statistic for detecting the changes between two Wishart distributed multilook covariance matrices.
Changes are finally detected by a decision threshold to the test statistic to distinguish change from no-change. In the present paper, the thresholding is performed using the generalized Kittler and Illingworth’s minimum-error algorithm (K&I for short) [10] on the SRW image. In the proposed method, the generalized Gamma distribution, denoted GΓD, is applied for modeling change and no-change classes in the SRW image. The GΓD was first proposed by Stacy [15] and has been widely applied in many fields, e.g., [14]. This distribution has a highly fixable form and good fitting capability to the histograms of change and no-change classes. Parameters of the probability density function (PDF) in this study are estimated using the method of log-cumulants (MoLC). This estimation method has been adopted in the analysis and processing of SAR images, e.g., [16]-[17]. The block diagram of the proposed unsupervised change detection method is presented in Fig. 1.