Conference Agenda

Quad-pol (QP) SAR systems acquire comprehensive polarimetric information by transmitting two and receiving four polarizations. However, they require higher pulse repetition rates (PRF) and more complex hardware, resulting in reduced swath width. Compact-pol (CP) systems, a specialized dual-pol mode, offer wider coverage with lower PRF requirements and have proven effective for land cover mapping, crop classification, soil moisture estimation, vegetation characterization, and ship detection.

Two main approaches exist for utilizing CP SAR data: extracting polarimetric features through decompositions, and reconstructing the 3×3 QP covariance matrix from the 2×2 CP covariance matrix. This work focuses on the latter approach, enabling application of QP processing techniques to reconstructed data.

Most reconstruction methods rely on Souyris' reflection symmetry assumption to relate CP and QP matrix elements, iteratively determining cross-polarization intensity. However, this assumption is valid primarily for volume-dominated scatterers and often fails for other scattering mechanisms.

We propose a deep learning method that reconstructs QP data from CP data without assuming reflection symmetry. The network uses real and imaginary components of the complex covariance matrix to directly map CP to QP data. The model was trained and evaluated on multiple PolSAR datasets to demonstrate performance across diverse scenarios.

Performance was assessed using Hotelling Lawley Trace (HLT), Determinant Ratio Test (DRT), and Wishart Distance for qualitative evaluation, alongside Mean Square Error (MSE), Mean Absolute Error (MAE), and Coherence Index (COI) for quantitative analysis. Results were compared against Souyris' and Nord's classical reconstruction methods.