Measuring the Earth’s above-ground biomass is a mission-critical requirement for understanding the global carbon cycle. While high-frequency sensors often saturate in dense forests, the ESA BIOMASS mission’s use of P-band SAR offers a breakthrough by providing the canopy penetration necessary to interact with large woody structures. However, this perspective is clouded by speckle noise a physical corruption that threatens the integrity of the scattering matrix. This work represents a specialized evolution of a polarization-agnostic methodology, transitioning a general purpose framework into a mission specific tool for P-band PolSAR. By adapting a shared convolutional backbone with a dynamic masking data loader, we enable a unified approach to despeckling that maintains the physical and statistical authenticity of the radar return. Training on real multi-temporal data and utilizing a spatio-temporal MuLog-filtered supervision strategy, we demonstrate that this evolved architecture effectively preserves the sub-canopy textures essential for carbon density estimation. The estimation of biomass is a critical variable in climate modeling, where the polarimetric signature serves as the primary signal for woody structures. Traditional filters often smooth forest textures, masking small-scale variations. Deep learning offers a solution, but adoption has been hindered by rigid architectures. Our approach builds upon an established polarization-agnostic framework, evolving its core logic to meet the rigorous physical demands of the BIOMASS mission. While BIOMASS data is primarily characterized by its Full-Pol (quad-polarization) nature, we retain the methodology’s underlying flexibility to ensure the network learns universal spatial manifolds that are not tethered to a single modality, thereby increasing the robustness of the feature extraction layers. The core innovation, inherited and refined for this application, is the ability to handle heterogeneous polarimetric inputs within a single unified architecture. Rather than using fixed-input dimensions, we implement a specialized masking-based data loader. In this approach, the network is designed to accept a maximum-capacity input tensor (e.g., a 9-channel stack representing the 3 x 3 covariance matrix C). During the data loading phase, a binary mask is applied to the input channels. For the BIOMASS mission's Full-Pol requirements, the mask acts as a gate that ensures all polarimetric relationships are preserved while the shared convolutional filters learn features that are invariant to channel presence. This masking strategy allows the model to leverage spatial textures found across the scattering matrix while remaining structurally flexible a direct evolution of previous work designed to process diverse mission modes without the need for costly retraining. One of the primary critiques of DL-based despeckling is the reliance on synthetic Gaussian noise for training. To address this, we leverage multi-temporal stacks from real data: Spatio-Temporal Averaging: We take N co-registered images of the same forest track. By averaging them in the temporal domain, we significantly reduce speckle while maintaining the spatial resolution of the original sensor. MuLog Integration: The resulting images are processed using the Multi-channel Logarithmic framework. This transforms the multiplicative Wishart noise into an additive Gaussian-like space, allowing the CNN to optimize its weights more effectively. The resulting "pseudo-clean" labels represent the true statistical behavior of the scattering, providing a much more rigorous training signal than synthetic noise models. A significant portion of our technical development focused on engineering a loss function L that transcends simple pixel-wise comparisons. In the context of P-band PolSAR, a standard Mean Squared Error often leads to over-smoothed images that strip away the texture essential for distinguishing forest types. To avoid this, we utilize a multi-objective optimization strategy centered on the Frobenius loss and the Gradient loss. In PolSAR, the physical properties of the target are encapsulated in the matrix structure; therefore, the network must minimize the distance between the estimated matrix and the pseudo-clean reference across all elements simultaneously. While the Frobenius loss handles the "physics" of the signal, the gradient loss handles the "geography" of the scene. P-band images are characterized by significant texture variations caused by the forest canopy’s uneven structure. To prevent the convolutional layers from blurring these vital edges, the gradient loss penalizes the difference between the spatial gradients of the denoised image and the reference. Ultimately, this framework provides the rigorous, scalable data processing needed for the BIOMASS mission ground segment. By evolving a polarization-agnostic model, we address the operational reality that while BIOMASS is a Full-Pol mission, the robust feature extraction granted by this methodology drastically reduces computational overhead and increases reliability. This approach demonstrates that we can clean the radar signal without destroying the delicate sub-canopy textures that signal carbon density. By bridging the gap between high-performance deep learning and the strict laws of radar physics, this evolved framework ensures that the BIOMASS mission provides the high-fidelity measurements needed to accurately track the pulse of our planet's forests.
Exploiting polarization-agnostic deep Learning method for PolSAR despeckling on recent BIOMASS data
Passarello G.
;Vitale S.;Ferraioli G.;Pascazio V.
2026-01-01
Abstract
Measuring the Earth’s above-ground biomass is a mission-critical requirement for understanding the global carbon cycle. While high-frequency sensors often saturate in dense forests, the ESA BIOMASS mission’s use of P-band SAR offers a breakthrough by providing the canopy penetration necessary to interact with large woody structures. However, this perspective is clouded by speckle noise a physical corruption that threatens the integrity of the scattering matrix. This work represents a specialized evolution of a polarization-agnostic methodology, transitioning a general purpose framework into a mission specific tool for P-band PolSAR. By adapting a shared convolutional backbone with a dynamic masking data loader, we enable a unified approach to despeckling that maintains the physical and statistical authenticity of the radar return. Training on real multi-temporal data and utilizing a spatio-temporal MuLog-filtered supervision strategy, we demonstrate that this evolved architecture effectively preserves the sub-canopy textures essential for carbon density estimation. The estimation of biomass is a critical variable in climate modeling, where the polarimetric signature serves as the primary signal for woody structures. Traditional filters often smooth forest textures, masking small-scale variations. Deep learning offers a solution, but adoption has been hindered by rigid architectures. Our approach builds upon an established polarization-agnostic framework, evolving its core logic to meet the rigorous physical demands of the BIOMASS mission. While BIOMASS data is primarily characterized by its Full-Pol (quad-polarization) nature, we retain the methodology’s underlying flexibility to ensure the network learns universal spatial manifolds that are not tethered to a single modality, thereby increasing the robustness of the feature extraction layers. The core innovation, inherited and refined for this application, is the ability to handle heterogeneous polarimetric inputs within a single unified architecture. Rather than using fixed-input dimensions, we implement a specialized masking-based data loader. In this approach, the network is designed to accept a maximum-capacity input tensor (e.g., a 9-channel stack representing the 3 x 3 covariance matrix C). During the data loading phase, a binary mask is applied to the input channels. For the BIOMASS mission's Full-Pol requirements, the mask acts as a gate that ensures all polarimetric relationships are preserved while the shared convolutional filters learn features that are invariant to channel presence. This masking strategy allows the model to leverage spatial textures found across the scattering matrix while remaining structurally flexible a direct evolution of previous work designed to process diverse mission modes without the need for costly retraining. One of the primary critiques of DL-based despeckling is the reliance on synthetic Gaussian noise for training. To address this, we leverage multi-temporal stacks from real data: Spatio-Temporal Averaging: We take N co-registered images of the same forest track. By averaging them in the temporal domain, we significantly reduce speckle while maintaining the spatial resolution of the original sensor. MuLog Integration: The resulting images are processed using the Multi-channel Logarithmic framework. This transforms the multiplicative Wishart noise into an additive Gaussian-like space, allowing the CNN to optimize its weights more effectively. The resulting "pseudo-clean" labels represent the true statistical behavior of the scattering, providing a much more rigorous training signal than synthetic noise models. A significant portion of our technical development focused on engineering a loss function L that transcends simple pixel-wise comparisons. In the context of P-band PolSAR, a standard Mean Squared Error often leads to over-smoothed images that strip away the texture essential for distinguishing forest types. To avoid this, we utilize a multi-objective optimization strategy centered on the Frobenius loss and the Gradient loss. In PolSAR, the physical properties of the target are encapsulated in the matrix structure; therefore, the network must minimize the distance between the estimated matrix and the pseudo-clean reference across all elements simultaneously. While the Frobenius loss handles the "physics" of the signal, the gradient loss handles the "geography" of the scene. P-band images are characterized by significant texture variations caused by the forest canopy’s uneven structure. To prevent the convolutional layers from blurring these vital edges, the gradient loss penalizes the difference between the spatial gradients of the denoised image and the reference. Ultimately, this framework provides the rigorous, scalable data processing needed for the BIOMASS mission ground segment. By evolving a polarization-agnostic model, we address the operational reality that while BIOMASS is a Full-Pol mission, the robust feature extraction granted by this methodology drastically reduces computational overhead and increases reliability. This approach demonstrates that we can clean the radar signal without destroying the delicate sub-canopy textures that signal carbon density. By bridging the gap between high-performance deep learning and the strict laws of radar physics, this evolved framework ensures that the BIOMASS mission provides the high-fidelity measurements needed to accurately track the pulse of our planet's forests.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
