Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction

Ning Ni, Libao ZhangCVPR 2025 · DOI

Satellite video has become an indispensable tool in modern remote sensing, facilitating critical applications such as disaster response, urban planning, and traffic monitoring. However, the utility of this data is often hampered by a "double-degradation" phenomenon: the inherent low-resolution (LR) nature of satellite sensors combined with atmospheric haze that obscures fine details.

In a recent study presented at CVPR, researchers Ning Ni and Libao Zhang from Beijing Normal University introduced ODPNet (Optimal Joint Degradation Pattern Network). This framework is designed to address both dehazing and "continuous-scale" super-resolution (SR) within a single, unified deep learning architecture.

The ODPNet Framework

Traditional approaches to this problem typically employ a "two-stage process" (TSP), where a video is first dehazed and then super-resolved. However, this sequential approach often leads to error propagation, where artifacts or inaccuracies from the dehazing stage are amplified during the super-resolution process.

ODPNet shifts this paradigm by treating dehazing and SR as a joint optimization problem. By modeling the degradation patterns simultaneously, the network can better preserve weak textures and recover fine-grained details that are often lost in sequential pipelines.

Comparison of the traditional two-stage process versus the proposed joint ODPNet model, with visual results.

Technical Architecture

The ODPNet architecture is composed of four primary modules designed to handle the specific complexities of satellite imagery, such as low contrast and complex motion:

  1. Prior-based Feature Soft Dehazing Module (PFSDM): Rather than relying solely on data-driven black-box methods, PFSDM integrates a physics-based atmospheric scattering model. It predicts specific dehazing factors ($K$ and $B$) to remove haze at the feature level, leading to more natural color restoration.
  2. Spatiotemporal Self-Attention (SSA): To ensure temporal consistency, the SSA module captures long-range dependencies across neighboring frames. This allows the model to leverage information from adjacent time steps to fill in missing spatial details in the target frame.
  3. Tri-branch Cross-aggregation Block (TCB): Satellite images often suffer from "weak textures." The TCB addresses this by using three parallel branches—spatial attention, channel attention, and interaction aggregation—to mine and enhance subtle details.
  4. Cross-scale Feature Top-k Selection Transformer (CFTST): This module enables "continuous-scale" reconstruction. Unlike standard models fixed to integer scales (e.g., 2x or 4x), the CFTST uses a Transformer to adaptively select the most relevant latent codes for any given coordinate, allowing for smooth magnification at arbitrary resolutions.
The complete ODPNet architecture, showcasing the integration of physics-based dehazing and continuous-scale super-resolution.

Experimental Evaluation

The researchers evaluated ODPNet using the Jilin-1 satellite video dataset, which features a ground resolution of approximately 1 meter. To simulate real-world conditions, the team applied synthetic haze and bicubic downsampling to the high-resolution source footage.

Key Results

  • Quantitative Superiority: ODPNet significantly outperformed state-of-the-art two-stage combinations. For instance, on the "Clip002" test set at an 8x scale, ODPNet achieved a PSNR of 26.26 dB, compared to 24.36 dB for the leading two-stage method.
  • Model Efficiency: Despite its advanced capabilities, ODPNet is remarkably efficient. It utilizes only 7.3M parameters, whereas many competing two-stage pipelines require between 15M and 25M parameters.
  • Visual Clarity: Qualitative results demonstrate that ODPNet effectively removes the "white veil" of haze, producing sharper edges and more accurate color reproduction than traditional methods.

Conclusion and Future Directions

ODPNet represents a significant advancement in satellite remote sensing by successfully unifying dehazing and continuous-scale super-resolution. By embedding physical priors into a deep learning framework, it achieves superior restoration quality with high parameter efficiency.

While the current model excels with uniform haze, future research may focus on adapting the architecture for non-homogeneous atmospheric conditions, such as variable cloud cover. Additionally, optimizing the Transformer components for real-time inference on satellite hardware remains a promising avenue for deployment in live monitoring systems.

Summary updated

BibTeX
@inproceedings{ni2025hazy,
  author = {Ning Ni and Libao Zhang},
  title = {{Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.01184},
}
Figure