Abstract

Reinforcement learning (RL) has emerged as one of the most promising and powerful techniques in deep learning. Despite its success, there have been very few practical applications of RL in computer vision tasks, where supervised learning is most dominant. In this thesis, we reformulate Road Extraction from Satellite Imagery as an RL problem. We aim to address some of the challenges of supervised learning methods and open the door for future works. To the best of our knowledge, this is the first time RL has been successfully applied to the complex real-world problem of road extraction, where the challenges are partially-observable, large-scale environments, and long time horizons. First, we design an environment, an action space, and a reward function that fully captures the problem as a partially observable Markov decision process. We propose a novel neural network architecture that captures the multi-modality in the input data. Then, we propose methods to address the challenges that come with the long time horizon aspect of the environment, and optimize and improve the policy efficiently. Due to the large-scale nature of the problem, we employ self-supervised representation learning to reduce the computational cost and increase the performance of the policy. We present experiments on satellite images of fifteen cities that demonstrate comparable performance to state-of-the-art methods.