Deep Learning optimization in remote sensing image segmentation using dilated convolutions and ShuffleNet

Semantic segmentation of land use land cover data using deep learning networks have gained significant importance in the remote sensing domain. However, deep learning architectures are computation-intensive. In this research, we propose an Atrous Shuffle-UNet network, which is designed to be lightweight. The network comprises of modified ShuffleNet units which are arranged in a similar network structure as the UNet. Atrous convolution in the proposed network increases the receptive field of the network enabling faster convergence. We compare the proposed network to state of the art deep learning architectures such as UNet, UNet with ResNet modules and a UNet with standard ShuffleNet modules. The proposed changes in the ShuffleNet units enable the network to outperform these architectures and do so with significantly less parameters.