Cloud Cover Removal from Remote Sensing Data using GANs Based on Attention Mechanism

Cloud interference is a significant challenge in remote sensing applications, impacting the quality and reliability of data used for environmental analysis, disaster monitoring, and urban planning. In this study, we present an innovative approach to remove clouds from remote sensing data by employing Generative Adversarial Networks (GANs) with an integrated attention mechanism. Our method is designed to effectively restore regions obscured by clouds while preserving the integrity of the remaining image. We introduce a specialized architecture called U-net based Attention GAN (UAGAN) to address cloud cover issues. In this architecture, the generator network utilizes attention mechanisms to identify and rectify cloud-affected areas. The attention mechanism dynamically emphasizes relevant spatial regions during cloud removal, enhancing the precision of the restoration process. To evaluate our approach, we conducted experiments using the RICE dataset and trained our UAGAN model. Our method’s effectiveness was assessed using quantitative metrics, including the Structural Similarity Index (SSIM), Root Mean Squared Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR). Comparative analysis of three different models, namely vanilla GAN, U-net GAN, and UAGAN, demonstrates that UAGAN outperforms the others with impressive results, achieving a PSNR of 29.901 dB, RMSE of 7.891, and SSIM of 0.971.This research contributes to the advancement of cloud removal techniques in remote sensing, offering a promising solution for improving data quality and enhancing the accuracy of downstream applications. The integration of GANs and attention mechanisms showcases the potential for innovative approaches to address complex challenges in remote sensing data analysis.