Reflection of and vision for the decomposition algorithm development and application in earth observation studies using PolSAR technique and data

Highlights

• A separation factor is studied to delineate an azimuthally symmetric or asymmetric target.

• A diplane model models an azimuthally asymmetric target in an urban area.

• A procedure derives an equivalent azimuth-orientation angle for the diplane model.

• An algorithm of balanced usability and procedural complexity to decompose PolSAR data is developed.

Abstract

After reflecting on the past decomposition studies using the polarimetric synthetic aperture radar (PolSAR) technique and data in Earth observation (EO) studies, three primary issues are identified. Elements C₁₂ and C₃₂ of a covariance matrix, [C], are essential in the decomposition and cannot be ignored. Existing algorithms cannot adequately distinguish urban targets with large azimuth orientation angles from vegetation. The algorithms are complex in the formulation and procedure. To resolve the issues and envision future algorithm development, we have articulated three key modules. They are a separation factor to separate an azimuthally symmetric or asymmetric target, a diplane to model an asymmetric target in an urban area, and a procedure to derive an equivalent azimuth-orientation angle for the diplane. Then, a four-component decomposition algorithm was developed. The algorithm has been applied to multiple airborne and spaceborne PolSAR C- and L-band datasets covering areas in Canada, France, Morocco, and the USA. The primary radar target types included trihedral and dihedral corner reflectors (CRs), airport runway/taxiway, urban targets with azimuth-orientation angles ranging between 0° and 45°, ocean and inland water surfaces, city parks, grassland, forests, farmland, and desert. The separation factor delineates a symmetric or asymmetric target at a correct average rate of 92.7%. The diplane coupled with the derived equivalent azimuth-orientation angles correctly modeled radar backscatter from dihedral CRs and urban asymmetric targets. The algorithm delineated each type of ground target with an appropriate and dominant single, double, or volume scattering. Furthermore, the algorithm has four readily interpretable components, and its mathematical expression is not complicated. Therefore, the primary objectives to resolve the above three issues and to have an algorithm with well-balanced usability in EO studies and complexity in formulation and procedure are achieved.

Introduction

The mono-state polarimetric synthetic aperture radar (PolSAR) characterizes a ground target by measuring the amplitude and phase from the incident radar wave's backscattering. Usually, the wave is linearly polarized horizontally (h) and vertically (v). Measured h and v elements are typically expressed in a scattering, covariance, or coherency matrix. Nevertheless, the interpretation of the matrices in Earth observation (EO) studies is not straightforward. As the understanding and use of the PolSAR technique, scattering mechanism, and data in forests are demanded, one develops complicated radar backscatter models. They include the MIMICS (McDonald and Ulaby, 1993; Ulaby et al., 1990), Santa Barbara microwave backscattering models (Sun et al., 1991; Wang et al., 1993; Wang et al., 1995; Wang et al., 1998), 3-D models (Liu et al., 2010; Sun and Ranson, 1995), models for pine forest (Lang et al., 1994), coherent scattering model for tree canopies based on the Monte Carlo simulation of scattering from fractal-generated trees (Lin and Sarabandi, 1999; Sarabandi and Lin, 2000), and multilayer-multispecies vegetation model (Burgin et al., 2011). Primary burdens in these forward modeling studies include the model parameterization and demand of in situ data.

Meanwhile, with the eigenvalue/eigenvector analysis, the PolSAR data were decomposed in forests to understand scattering mechanisms and applications in forested environments (van Zyl, 1993; Wang and Davis, 1997). Cloude and Pottier (1996) reviewed and unified three types of decomposition algorithms. Freeman and Durden (1998) developed a three-component decomposition algorithm for forests. These developments are valued as viable alternatives to the forward modeling approaches in understanding radar backscatter in forests because of their simplicity in the formulation and no demand for the in situ data. When the decomposition algorithms primarily developed for forests are extended to other land cover types such as an urban area, their deficiency is found. Due to strong volume scattering from urban targets such as buildings with large azimuth-orientation angles, the algorithms cannot separate the volume scattering from tree canopies or urban targets. The phase information of co- and cross-pol cross-products was studied for the separation (Atwood and Thirion-Lefevre, 2018; Gan et al., 2019). With the subaperture approach (Ling et al., 2021) and the correlation analysis of cross-product of two cross-pol datasets (Zhang et al., 2021), the delineation of a stationarity target (e.g., a building) and a non-stationarity target (e.g., a tree canopy) were investigated.

The inseparability is also one fundamental impetus to further the decomposition algorithm development by adding new components or revising existing ones. Yamaguchi et al. (2005) added the helix to the Freeman-Durden algorithm (Freeman and Durden, 1998), and Xiang et al. (2015) proposed a five-component algorithm. Singh and Yamaguchi (2018) and Singh et al. (2019) included additional ones. Moreover, the azimuth-angle deorientation (e.g., Lee and Ainsworth, 2011) was included to reduce the volume scattering from urban targets with large azimuth orientation angles (Susaki et al., 2014; Yamaguchi et al., 2011). Numerous revisions of existing algorithm components were done and summarized (e.g., Chen et al., 2014; Duan and Wang, 2017; Li and Zhang, 2018). Furthermore, Du and Wang (2016) employed interferometric coherence to separate vegetation from an urban area. The smallest eigenvalue of the three eigenvalues representing single, double, and volume scattering components was used to identify the decomposed volume scattering (An and Lin, 2019). Although the current improved and revised algorithms are more powerful now in EO studies, urban targets with large azimuth orientation angles still cannot be correctly identified. Moreover, the revision of existing components and the addition of new ones make the component formulations and decomposition algorithms more complex procedurally.

In light of the increased decomposition algorithms' complicity, a reconsideration was argued (Wang et al., 2019). Dey et al. (2020) recharacterized three scattering components using the roll-invariant scattering-type parameters, and Ratha et al. (2020) explored the factorization of scattering power and unsupervised classification scheme. A search with “decomposition” as a keyword at the 2020 IEEE International GeoScience and Remote Sensing Symposium (https://igarss2020.org/technical_program.php), virtually held between September 26 and October 2, 2020, thirteen PolSAR decomposition studies are shown. Additionally, multiple civilian and operational airborne and spaceborne PolSARs exist worldwide (https://eoportal.org/web/eoportal/). No doubt, studying PolSAR decomposition algorithms will continue. Thus, it is time to reflect on how the EO research and application's advancement using the PolSAR technique and decomposition algorithm has been accomplished. What are the critical issues currently? Is it possible to propose a vision for future algorithm development to preserve the simplicity in the algorithm's formulation and procedure and the straightforward interpretation of the algorithm's components and outcomes? Thus, our study is framed with a reflection on the past and a vision for the future. The objective is to have an algorithm with well-balanced usability and complexity in formulation and procedure.