Spatial-frequency fusion network with learnable fractional fourier transform for remote sensing imaging enhancement

Abstract

Atmospheric haze significantly degrades the quality of remote sensing images, reducing visibility, distorting spectral information, and impairing downstream tasks such as land cover classification and infrastructure layout analysis. To overcome these challenges, this article proposes a novel spatial–frequency fusion network (termed SFFNet) with a learnable fractional Fourier transform for efficient remote sensing imaging enhancement. In the spatial domain, the SFFNet uses a multiscale spatial pyramid pooling block to capture both fine-grained details and global contextual information, while residual connections ensure robust feature learning and spatial detail preservation. In the frequency domain, a self-learned fractional Fourier transform module adaptively extracts haze-relevant features, leveraging a learnable parameter to dynamically adjust the fractional order of the transform. Furthermore, an attentive frequency gate selectively emphasizes critical frequency features based on the local features of the input image. To effectively address the challenges of nonuniform haze distribution, a self-attention-guided fusion mechanism is introduced, synergistically integrating spatial and frequency information. In addition, a hierarchical feature fusion strategy progressively refines multiscale features throughout the dehazing process, ensuring comprehensive and accurate haze removal. Experimental results on both synthetic and real-world remote sensing datasets show that the SFFNet achieves significant improvements in quantitative metrics and visual quality. Moreover, the SFFNet demonstrates strong practical potential in remote sensing object detection by improving accuracy and robustness.