Conditional Adversarial Networks Assisted Road Extraction in Remote Sensing Imagery
Download as PDF
Jianhua Li, Hongbing Ma
In this paper, we narrow down the task for conditional adversarial networks focusing on solving the binary segmentation problem for road extraction on remote sensing imagery. We constrain the objective of the generator of a conditional adversarial network to make it suitable to produce binary road predictions in an effective way. We evaluate our approach on the SpaceNet Roads dataset and our method shows promising results compared to standard segmentation models. Our best pixel level precision score on test set is 76.9%. This paper shows that adversarial networks with detailed images as input and binary mask as output can be optimized with certain adaption and optimization tweaks, showing potentials in solving computer vision problems.
Generative adversarial networks, road extraction, remote sensing image, machine learning