Deep Convolutional Encoder–Decoder Models for Road Extraction from Aerial Imagery

Road extraction from aerial imagery is not a trivial task. It plays a pivotal role in urban planning, navigation, disaster assessment and various other fields. It poses challenges due to complex scenarios and factors, including occlusion. Hence conventional methods prove to be inefficient for the purpose. Image segmentation and deep learning models are extensively employed in recent times to extract objects from images. In this paper, the performance of Unet architecture-based model has been improved by Resnet50, VGG16, DenseNet169, Xception and Efficientnet-b4. Further, to investigate the performance of Unet model, three other models FPN, PSPNet and PAN were implemented and evaluated on Massachusetts road dataset. The work presents the comparative analyses of the performance of models.