Cloud Detection and Removal from RGB Images using U-Net Semantic Segmentation and CloudGAN Models

Hussein, Aziza I, Hassan, Mennatall Essam, Ekpo, Sunday ORCID: https://orcid.org/0000-0001-9219-3759, Alyami, Ghadah S, Elias, Fanuel, Salah, Ibrahim and Mabrook, M Mourad (2024) Cloud Detection and Removal from RGB Images using U-Net Semantic Segmentation and CloudGAN Models. In: International Adaptive and Sustainable Science, Engineering and Technology (ASSET) Conference 2024, 16 July 2024 - 18 July 2024, Manchester. (In Press)

Accepted Version File not available for download. Available under License In Copyright. Download (533kB)

Abstract

Satellite-based imagery provides an indispensable tool for many applications, ranging from environmental surveillance to urban development and managing natural disasters. Nevertheless, the presence of clouds can often impede the useful-ness of these images by veiling significant details. In the current study, we proposed an innovative strategy for identify-ing and eliminating clouds within RGB satellite images employing deep learning techniques. This involves using a Cloud Generative Adversarial Network (CloudGAN) to carry out image inpainting tasks and U-Net for semantic segmentation. The proposed methodology yields encouraging outcomes, showcasing its ability to discern and eradicate clouds effective-ly, thereby enhancing the clarity and practicality of satellite imagery. The proposed approach demonstrates superior cloud removal compared to traditional methods, achieving a remarkable overall accuracy of 95\% in both cloud detec-tion and removal. This underscores its effectiveness in enhancing image quality and utility. The qualitative assessment confirms the models' ability to produce high-quality, cloud-free images, preserving essential features and details faithfully. Additionally, the inpainted images closely resemble the ground truth, affirming the accuracy of the models in cloud removal.