An Ensemble Wasserstein Generative Adversarial Network Method for Road Extraction From High Resolution Remote Sensing Images in Rural Areas

Road extraction from high resolution remote sensing (HR-RS) images is an important yet challenging computer vision task. In this study, we propose an ensemble Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) method called E-WGAN-GP for road extraction from HR-RS images in rural areas. The WGAN-GP model modifies the standard GANs with Wasserstein distance and gradient penalty. We add a spatial penalty term in the loss function of the WGAN-GP model to solve the class imbalance problem typically in road extraction. Parameter experiments are undertaken to determine the best spatial penalty and the weight term in the new loss function based on GaoFen-2 dataset. In addition, we execute an ensemble strategy in which we first train two WGAN-GP models using the U-Net and BiSeNet as generator respectively, and then intersect their inferred outputs to yield better road vectors. We train our new model with GaoFen-2 HR-RS images in rural areas from China and also the DeepGlobe Road Extraction dataset. Compared with the U-Net, BiSeNet, D-LinkNet and WGAN-GP methods without ensemble, our new method makes a good trade-off between precision and recall with F1-score = 0.85 and IoU = 0.73.