Cloud Image Masking on Satellite Images Using U-Net Architecture of Deep Learning*

Satellite photos are commonly utilized for many activities such as surveys, map creation, and surveillance, and cloud detection can make using such high-quality images easier. Global yearly mean cloud cover is roughly 66 percent, according to the International Satellite Cloud Climatology. Clouds frequently obscure surfaces and objects, making it harder to extract usable information from photographs. 38-Cloud [1] is the dataset that is utilized in this investigation. In order to segment clouds in satellite images, we used only three input channels from the dataset: Red, Green, Blur, and NIR (near infrared). This prompted us to take an entirely different strategy and create our own U-Net from the ground up and not use the vision libraries for preparing the dataset. The implemented U-Net model is showing 96.59% accuracy. This paper also includes the recent art of work carried out in the Cloud detection application.