Perceptual dehazing of remote sensing images using global attention and Laplacian-Guided GANs for environmental applications
Anas M. Ali, Bilel Benjdira, Wadii Boulila · Ecological Informatics · 2025
Atmospheric haze significantly degrades the visual quality of remote sensing imagery by obscuring fine structural and textural details, thereby reducing the reliability of ecological and environmental assessments. This study presents GLADE-Net ( Global-Attention and Laplacian-Enhanced Dehazing Network ), a novel dual-stage deep learning architecture developed to improve the interpretability of high-resolution remote sensing data for environmental and ecological analysis. In the first stage, a global attention-driven encoder–decoder module processes non-overlapping grid patches to capture broad contextual dependencies and effectively suppress atmospheric haze. The second stage further refines the restored output through a dual-branch generative design that jointly operates in the spatial (RGB) and frequency (Laplacian) domains, enabling the recovery of both low-frequency structures and high-frequency details essential for accurate environmental interpretation. Comprehensive experiments conducted on the SateHaze1k and RICE benchmark datasets confirm that GLADE-Net surpasses contemporary state-of-the-art approaches in terms of both structural fidelity (PSNR, SSIM) and perceptual realism (LPIPS). Specifically, on the SateHaze1k dataset, it attains an average PSNR of 26.74 dB and an LPIPS of 0.0471, yielding improvements of 5.87% and 21.6% over the strongest baselines. On RICE1 and RICE2, the proposed model achieves PSNR scores of 36.56 dB and 34.27 dB, and LPIPS values of 0.0176 and 0.0746, respectively, demonstrating consistent gains across all metrics. These outcomes highlight the capability of GLADE-Net as a robust and practical framework for enhancing remote sensing imagery in ecological and atmospheric studies. All implementation materials, including datasets, source code, model weights, and evaluation scripts, are publicly accessible at: https://github.com/riotu-lab/GLADE-Net .