The aim of this project is to concieve a classifier capable of discerning the roads from background elements using a small dataset of satellite images originating from Google maps. Both standard convolutional neural networks and U-net variants were implemented and tested. Our best result was achieved using an Attention U-net with extended data augmentation and hyperparameter tuning yielding a F1 score of 0.85 on a hidden test set from AIcrowd.
This repo contains the data, code and environemnt for reproductibility of our results as well as the full report.