Road segmentation in satellite images using deep learning

We demonstrate several approaches for dealing with road segmentation problems. Some rely on the encoder-decoder method, some on the Generative Adversarial Network (GAN) technique, and others on a fully convolutional network. The approaches utilizing Encoder- Decoder and GAN seem to have promise. Due to the great performance of Encoder-Decoder Deep Convolutional Neural Networks in many segmentation problems. Our aim is to apply all recent model architectures that use the DCEP network as a primary base model on two open-source data sets DeepGlope, and Massachusetts. We choose the most common encoder-decoder models that proved great performance for different data sets of image segmentation. We choose Unet, FPN, PSPNet, Unet++, PAN, LinkNet, DeepLab- v3, DeepLab-v3+, and MA-Net for our experiments and we give a brief comparison based on the result. We show the results for each model we use, both with and without the bilateral filter, and we show how the IOU (Intersection Over Union) and Dice loss of the results for all models on the Massachusetts data set are very similar. In an effort to improve model performance, we also use different data augmentation parameters, however, the results are the same for this data set. The Unet model has an excellent IOU for the DeepGlobe data set, scoring 95.46% accuracy.