GATE-Dehaze: A GAN-attention-transformer ensemble for satellite image dehazing
Ishaan Gupta, Yesha Rawal, Shiwani Gupta, Sujata Alegavi · Turkish Journal of Remote Sensing · 2026
The Satellite imagery serves as a cornerstone for various human endeavours including environmental monitoring, urban planning, and disaster management. However, the efficacy of these images heavily relies on their quality, with atmospheric phenomena like haze posing significant challenges to detailed analysis. Addressing this issue is crucial for extracting meaningful insights. This research undertakes the task of analyzing current methodologies and proposes leveraging a new algorithm GATE-Dehaze, GAN-Attention-Transformer-Ensemble for dehazing, which consists of an adaptive ensemble algorithm applied on the output of four GAN based models that includes: a standard U-Net architecture, a Transformer-enhanced U-Net, a Transformer U-Net with CBAM integration and a Transformer U-Net with channel attention. GATE-Dehaze is used here for image-to-image translation, facilitating the conversion of source images (which are hazy) into clearer (dehazed) target images. The evaluation of our method is conducted using metrics like Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) scores. On benchmarking datasets like SateHaze1k with dense haze, our model performed better than other SOTA models. Our proposed algorithm increased the SSIM score between ground truth and dehazed image from 0.9061 to 0.9257 on thin haze and 0.9264 to 0.9516 on moderate haze. By employing our approach, the aim is to generate dehazed satellite images that enable more precise analysis of the Earth's surface across varying spatial and temporal scales. Ultimately, the availability of high-quality, dehazed satellite imagery holds the potential to unlock invaluable insights that can inform decision-making processes in diverse fields ranging from environmental conservation to urban development and disaster response.