DPFNet: A Detail-Preserved Pyramid Fusion Network for Remote Sensing Image Dehazing
Shiyu Quan, Qunming Wang · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Remote sensing image dehazing is an important preprocessing step for improving the quality of optical imagery and ensuring reliable downstream applications. Most of existing (e.g., deep learning-based) approaches were developed based on synthetic hazy data (e.g., trained using simulated hazy images), and their effectiveness often drops on real remote sensing images due to complex atmospheric variations and sensor-specific haze responses. This leads to loss of spatial details, spectral distortions, and limited adaptability to scenes with diverse haze densities. To overcome these challenges, we proposed a Detail-Preserved Pyramid Fusion Network (DPFNet), which was designed directly for real-world hazy images (trained using real hazy images). It enhances dehazing through three key components: 1) Haze sensing fusion mechanism. Based on spatial attention guidance and channel attention enhancement, the haze-related spatial response maps and multi-scale features were fused adaptively, thereby enhancing its robustness under diverse haze conditions; 2) High-frequency detail preservation module. High-pass filtering together with residual channel attention was employed to preserve high-frequency details and reduce the information loss induced by over-smoothing; 3) Composite loss function. A joint loss combining pixel-level reconstruction, multi-scale consistency, and structural similarity constraints was used to maintain spatial structural fidelity and spectral consistency in the dehazed results. Experiments on real-world remote sensing datasets show that DPFNet consistently outperforms seven representative benchmark methods across various haze densities, achieving greater preservation of spatial structure and spectral property.