Mixer U-Net: An Improved Automatic Road Extraction from UAV Imagery

Abstract

Automatic road extraction from unmanned aerial vehicle (UAV) imagery has been one of the major research topics in the area of remote sensing analysis due to its importance in a wide range of applications such as urban planning, road monitoring, intelligent transportation systems, and automatic road navigation. Thanks to the recent advances in Deep Learning (DL), the tedious manual segmentation of roads can be automated. However, the majority of these models are computationally heavy and, thus, are not suitable for UAV remote-sensing tasks with limited resources. To alleviate this bottleneck, we propose two lightweight models based on depthwise separable convolutions and ConvMixer inception block. Both models take the advantage of computational efficiency of depthwise separable convolutions and multi-scale processing of inception module and combine them in an encoder–decoder architecture of U-Net. Specifically, we substitute standard convolution layers used in U-Net for ConvMixer layers. Furthermore, in order to learn images on different scales, we apply ConvMixer layer into Inception module. Finally, we incorporate pathway networks along the skip connections to minimize the semantic gap between encoder and decoder. In order to validate the performance and effectiveness of the models, we adopt Massachusetts roads dataset. One incarnation of our models is able to beat the U-Net’s performance with 10× fewer parameters, and DeepLabV3’s performance with 12× fewer parameters in terms of mean intersection over union (mIoU) metric. For further validation, we have compared our models against four baselines in total and used additional metrics such as precision (P), recall (R), and F1 score.