Enhancing Cloud Detection Performance: A Comparative Study of CNN Models and Architectures

Clouds play a pivotal role in shaping Earth's climate system and significantly influence climate change dynamics. Accurately monitoring cloud cover changes is essential for understanding broader climate patterns. However, detecting clouds through satellite remote sensing encounters challenges like data noise and environmental factors such as ice and snow. Leveraging advances in artificial intelligence, particularly deep learning techniques, shows promising potential for addressing these hurdles. In this study, various convolutional neural networks (CNN) architectures, including VGG-19, VGG-16, ResNet-51, ResNet101, and DenseNet-121, were assessed for their ability to extract features These CNN models were combined with UNet to generate masks capable of distinguishing cloud (foreground) from non-cloud (background) areas. The results demonstrated a significant improvement in accuracy compared to conventional cloud detection methods. The combination of DenseNet-121 and UNet achieved the highest accuracy, reaching 98.69%. This approach not only promises to enhance cloud detection accuracy but also holds potential for improving optical image processing tasks more broadly.