PhyDAE: Physics-Guided Degradation-Adaptive Experts for All-in-One Remote Sensing Image Restoration

TL;DR

PhyDAE tackles the problem of restoring remote sensing imagery under complex, multi-type degradations by integrating explicit physics priors into a two-stage, cascaded all-in-one framework. It introduces Residual Manifold Projector (RMP) and Frequency-Aware Degradation Decomposer (FADD) to mine degradation information from residuals, feeding a physics-aware mixture-of-experts whose routing is guided by temperature-controlled Top-K sparse activation. The restoration objective combines a degradation-aware optimal transport term with frequency-domain priors and multi-term losses (pixel, balance, contrast) to ensure physical consistency, stability, and discriminative degradation handling. Empirically, PhyDAE achieves state-of-the-art or competitive results on MD-RSID, MD-RRSHID, and MDRS-Landsat across dehazing, denoising, deblurring, and low-light tasks, while maintaining superior efficiency (17.21M parameters, 71.63 GFLOPs). This work enables effective, physics-consistent remote sensing restoration with scalable deployment potential in real-world scenarios.

Abstract

Remote sensing images inevitably suffer from various degradation factors during acquisition, including atmospheric interference, sensor limitations, and imaging conditions. These complex and heterogeneous degradations pose severe challenges to image quality and downstream interpretation tasks. Addressing limitations of existing all-in-one restoration methods that overly rely on implicit feature representations and lack explicit modeling of degradation physics, this paper proposes Physics-Guided Degradation-Adaptive Experts (PhyDAE). The method employs a two-stage cascaded architecture transforming degradation information from implicit features into explicit decision signals, enabling precise identification and differentiated processing of multiple heterogeneous degradations including haze, noise, blur, and low-light conditions. The model incorporates progressive degradation mining and exploitation mechanisms, where the Residual Manifold Projector (RMP) and Frequency-Aware Degradation Decomposer (FADD) comprehensively analyze degradation characteristics from manifold geometry and frequency perspectives. Physics-aware expert modules and temperature-controlled sparse activation strategies are introduced to enhance computational efficiency while ensuring imaging physics consistency. Extensive experiments on three benchmark datasets (MD-RSID, MD-RRSHID, and MDRS-Landsat) demonstrate that PhyDAE achieves superior performance across all four restoration tasks, comprehensively outperforming state-of-the-art methods. Notably, PhyDAE substantially improves restoration quality while achieving significant reductions in parameter count and computational complexity, resulting in remarkable efficiency gains compared to mainstream approaches and achieving optimal balance between performance and efficiency. Code is available at https://github.com/HIT-SIRS/PhyDAE.

Figures (7)

• Figure 1: Overall architecture of the proposed PhyDAE framework for all-in-one remote sensing image restoration. The two-stage cascaded pipeline progressively transforms degradation information from implicit features to explicit decision signals through the integration of Frequency-Aware Degradation Decomposer (FADD) and Residual Manifold Projector (RMP) modules.
• Figure 2: Detailed architecture of the Residual Manifold Projector (RMP) module. The hierarchical structure incorporates Multi-Depthwise Convolution Head Transposed Attention (MDTA) and Gated Depthwise Convolution Feed-Forward Network (GDFN) to extract multi-scale residual embeddings from manifold geometry perspective.
• Figure 3: Internal structure of the Frequency-Aware Degradation Decomposer (FADD) module. The multi-scale frequency decomposition strategy employs convolutional kernels with varying receptive fields to capture degradation-specific spectral signatures for posterior probability estimation.
• Figure 4: Physics-aware expert networks and temperature-controlled sparse activation mechanism. (a) Dehazing expert incorporating atmospheric scattering model. (b) Denoising expert with spatial-adaptive filtering strategy. (c) Deblurring expert utilizing anisotropic Gaussian modeling. (d) Probabilistic expert allocation with Top-K sparse routing. (e) Low-light enhancement expert based on Retinex theory.
• Figure 5: Visual comparison of different all-in-one restoration methods on the MD-RSID dataset. The zoomed-in results are provided to highlight the restoration performance advantages of PhyDAE over state-of-the-art methods.