A unified deep learning Model for despeckling in multi-modal polarimetric SAR data

Polarimetric Synthetic Aperture Radar (PoISAR) imaging is a powerful tool for studying how surfaces and objects interact with electromagnetic waves. It allows detailed observation of land cover, vegetation, and human-made structures. However, because SAR data are acquired through a coherent imaging process, they are affected by speckle noisea multiplicative effect that reduces image clarity and makes physical or statistical analysis less reliable. Traditional methods for reducing speckle, such as multi-looking or model-based filters like Lee, Frost, and refined Gamma-MAP, rely on local statistics of pixel intensity or covariance matrices. While these approaches can effectively smooth noise and improve radiometric consistency, they often blur fine spatial details. Recent state-of-the-art despeckling methods have introduced several deep learning (DL) and data-driven approaches that have proven powerful in mitigating noise in SAR images while preserving structural details and avoiding blurring effects. However, most of these DL-based methods are designed for aspecific number of polarization channels-either single, dual, or fully polarimetric data. Therefore, there remains a need for new methods capable of handling SAR image denoising across different polarimetric modalities. To this end, we introduce a deep learning based framework for speckle suppression that works with any number of polarization channels. The key idea is a band-agnostic neural network architecture capable of handling single-, dual-, or quad-polarization data without any modification or retraining. The model uses ashared convolutional backbone to learn common spatial features across channels, while a cross-polarization attention mechanism captures the relationships between them preserving the physical information encoded in the scattering matrix. Our method is trained entirely on real PolSAR observations, where the reference (noise-free) images are generated using a spatio-temporal averaging strategy applied to time-series data. Multiple co-registered SAR images of the same area are temporally averaged and subsequently processed with the MuLog spatial filtering technique to produce a noise-suppressed dataset. These filtered images serve as pseudo-clean references, enabling the network to learn directly from real data while preserving the physical characteristics and statistical authenticity of SAR observations. During training, the loss function balances two goals: accurate spatial reconstruction and preservation of polarization coherence. This ensures that the denoised images stay true to the original scattering behavior, distinguishing random speckle variations from meaningful structural features. In summary, this work presents a polarization-agnostic, data-driven deep leaming approach to speckle reduction in PolSAR imagery. By combining multitemporal MuLog-filtered supervision with a flexible neural architecture, our method unites the strengths of traditional statistical filters and modern deep learning. The result is a practical, scalable, and physically consistent solution for improving the quality and interpretability of polarimetric SAR data across a wide range of remote sensing applications.