Remote Sensing Road Segmentation Based on Feature Extraction Optimization and Skeleton Detection Optimization

High-resolution remote sensing image segmentation is a mature application in many industrial-level image applications, such as those in the military, civil, and other fields. Scene analysis needs to be automated in high-resolution remote sensing images as much as possible. Nowadays, with the rise of deep learning algorithms, remote sensing image processing algorithms have made tremendous progress. Deep learning algorithms process unlabeled data by learning a certain amount of labeled data. We conducted a specific study on the road target with GF1 data collected in China, and the remote sensing image’s resolution was 2 m (Jin et al. in J Anhui Agri Sci 43:358–362, 2015 [Jin et al. in J Anhui Agri Sci 43:358–362, 2015]). According to the observation of road features in remote sensing images, it still has a large number of small roads that are difficult to distinguish in the 2 m resolution GF1 remote sensing image. Due to the limitations of the downsampling calculation of the fully convolutional neural network, it is easy to lose a lot of information on small roads (Long et al. in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3431–3440, 2015 [Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp 3431–3440]). Therefore, we have adjusted the feature extraction and backbone networ. We adopted EfficientNet (Tan and Le in Efficientnet: Rethinking model scaling for convolutional neural networks, 2019 [Tan M, Le QV (2019) Efficientnet: rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946]) as the skeleton network of the algorithm, and combined D-linkNet (Zhou et al. in CVPR workshops, pp. 182–186, 2018 [Zhou L, Zhang C, Wu M (2018) D-LinkNet: LinkNet with pretrained encoder and dilated convolution for high resolution satellite imagery road extraction. In: CVPR workshops, pp 182–186]) as our After preliminary training to proposed D2-LinkNet, based on the road samples, we can get very good prediction results. According to the existing prediction results, there is still a certain difference in the fitting of our segmentation results to the groundtruth. To solve this problem, based on the extraction and analysis of the skeleton of the road prediction results and the integration of different prediction results, we proposed a skeleton optimization algorithm to optimize our prediction results for road samples. And on the one hand, it complements the segmentation of small roads. On the other hand, the road segmentation result fits the road boundary more closely. The final experiment results show that the optimization algorithm effectively improves the prediction accuracy of the network, and we have achieved better results compared with D-LinkNet in our large range remote sensing dataset.