Computer Vision Project, NYU Fall 2022
In this project, we implemented multiple U-Net models to perform semantic segmentation of landcover from satellite images. Landcover.ai dataset was used to train three different U-Net models with different encoders to generate segmentation masks. We also performed data augmentation using various transforms to improve on classification accuracy and reduce overfitting. The performance of the trained models were evaluated and compared using various metrics such as Jaccard Index, Precision, Recall, F1 Score and Exact Match Ratio.