This project uses a custom made dataset created using satellite images from CARTOSAT and labelling them using an open source tool "labelme" link - https://github.com/wkentaro/labelme
The dataset was trained on a custom made U-Net model and the model achieved an accuracy of 95% on the validation split.