The objective of this project is to detect rooftops using deep learning. We have 180 satellite images from 5 regions in United States. We tried several model architectures such as Vgg19 Unet, Unet, Deep Lab, U2net to find the best model segmentation results. We then evaluate the results using both dice loss, jacaard loss and cross entropy loss. We then deploy the solution using Streamlit as web application, allowing user to upload their satellite image, or select map location by integrating Google Map API service.