Satellite imagery analysis for road segmentation using U-Net architecture

Road network plays a significant role in today's urban development. These are a vital part of applications such as automatic road navigation, traffic management, route optimization, etc. In this paper, we aim to explore the potential and performance of convolution neural network architecture, U-Net, performing semantic segmentation to detect roads. The model has been evaluated on Massachusetts Roads Dataset with the estimation of numerous parameters such as filter stride, learning rate, training epochs, data size, and various augmentation techniques. The employment of the proposed technique and hyperparameters to the U-Net architecture gives a boost to test accuracy and improves the dice coefficient by a value of 0.007. Experimental results thereby demonstrate that our model is computationally efficient and achieves comparable segmentation results.