Comparing ASM and Learning-Based Methods for Satellite Image Dehazing

Abstrak

Recent advancements in optical satellite technologies have significantly improved image resolution, providing more detailed information about Earth's surface. However, atmospheric interference, such as haze, is still a major factor in image capture. The interference results in visibility degradation of the acquired images, hindering computer vision tasks. Numerous studies have proposed various methods to recover haze-affected regions in satellite images, highlighting the need for more effective solutions. Motivated by this, this paper compares different atmospheric dehazing methods, including Atmospheric Scattering Model (ASM)-based and deep learning-based. The results show that SRD is the best ASM-based method, with a PSNR value of 19.09 dB and an SSIM of 0.908. Among deep learning models, DW-GAN achieves the best restoration results with a PSNR value of 26.22 dB and an SSIM of 0.959. SRD offers faster inference times, but still suffers from residual haze and noticeable color degradation compared to DW-GAN. In contrast, DW-GAN provides a more complete haze removal at the cost of higher computational demands than ASM-based methods.